Professional Documents
Culture Documents
V. E. Saouma - Finite Element II, Solid Mechanics PDF
V. E. Saouma - Finite Element II, Solid Mechanics PDF
FINITE ELEMENT II
Solid Mechanics
CVEN 6525
c VICTOR E. SAOUMA,
SPRING 2001
Contents
1 PREREQUISITE 11
1.1 Variational Formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 Finite Element Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.1 Strain Displacement Relations . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.1.1 Axial Members . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.1.2 Flexural Members . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.2 Virtual Displacement and Strains . . . . . . . . . . . . . . . . . . . . . . . 15
1.2.3 Element Stiness Matrix Formulation . . . . . . . . . . . . . . . . . . . . 15
1.2.3.1 Stress Recovery . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3 Direct Stiness Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.1 Global Stiness Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.1.1 Structural Stiness Matrix . . . . . . . . . . . . . . . . . . . . . 17
1.3.1.2 Augmented Stiness Matrix . . . . . . . . . . . . . . . . . . . . 17
1.3.2 Logistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.3.2.1 Boundary Conditions, [ID] Matrix . . . . . . . . . . . . . . . . . 18
1.3.2.2 LM Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
1.3.2.3 Assembly of Global Stiness Matrix . . . . . . . . . . . . . . . . 110
E 1-1 Assembly of the Global Stiness Matrix . . . . . . . . . . . . . . . . . . . 110
1.3.2.4 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
E 1-2 Direct Stiness Analysis of a Truss . . . . . . . . . . . . . . . . . . . . . . 113
E 1-3 Analysis of a Frame with MATLAB . . . . . . . . . . . . . . . . . . . . . 118
E 1-4 Analysis of a simple Beam with Initial Displacements . . . . . . . . . . . 120
2 INTRODUCTION 21
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 Elliptic, Parabolic and Hyperbolic Equations . . . . . . . . . . . . . . . . . . . . 21
E 2-1 Seepage Problem;(Bathe 1996) . . . . . . . . . . . . . . . . . . . . . . . . 23
E 2-2 Diusion Problem; (Bathe 1996) . . . . . . . . . . . . . . . . . . . . . . . 25
E 2-3 Wave Equation, (Bathe 1996) . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3 Solution of Discrete-System Mathematical models . . . . . . . . . . . . . . . . . . 28
2.3.1 Steady State Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.1.1 Elastic Spring . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3.1.2 Heat Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
2.3.1.3 Hydraulic Network . . . . . . . . . . . . . . . . . . . . . . . . . . 211
2.3.1.4 DC Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
2.3.2 Equivalent Truss/Direct Stiness Models . . . . . . . . . . . . . . . . . 213
2.3.2.1 Nonlinear Elastic Spring . . . . . . . . . . . . . . . . . . . . . . 215
Draft
02
. . . . . . . 215
2.3.3.1 Dynamic Elastic System . . . . . . . . . . . . . . . . . . . . . . 216
2.3.3.2 Transient Heat Flow . . . . . . . . . . . . . . . . . . . . . . . . . 216
2.3.4 Eigenvalue Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
2.3.4.1 Free Vibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
2.3.4.2 Column Buckling . . . . . . . . . . . . . . . . . . . . . . . . . . 218
2.4 Solution Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
2.4.1 Euler Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
E 2-4 Flexure of a Beam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
2.5 Computer Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
2.6 Examples of applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
3 FUNDAMENTAL RELATIONS 31
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.1.1 Indicial Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.1.2 Tensor Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.1.3 Voigt Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.2 Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Vector Fields; Solid Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2.1 Kinetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2.1.1 Force, Traction and Stress Vectors . . . . . . . . . . . . . . . . . 35
3.2.1.2 Traction on an Arbitrary Plane; Cauchys Stress Tensor . . . . . 36
E 3-1 Stress Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.2 Kinematic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.3 Fundamental Laws of Continuum Mechanics . . . . . . . . . . . . . . . . 310
3.2.3.1 Conservation of Mass; Continuity Equation . . . . . . . . . . . . 312
3.2.3.2 Linear Momentum Principle; Equation of Motion . . . . . . . . 312
3.2.3.3 Conservation of Energy; First Principle of Thermodynamics . . 313
3.2.4 Constitutive Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
3.2.4.1 General 3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
3.2.4.2 Transversly Isotropic Case . . . . . . . . . . . . . . . . . . . . . 315
3.2.4.3 Special 2D Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
3.2.4.3.1 Plane Strain . . . . . . . . . . . . . . . . . . . . . . . . 316
3.2.4.3.2 Axisymmetry . . . . . . . . . . . . . . . . . . . . . . . . 316
3.2.4.3.3 Plane Stress . . . . . . . . . . . . . . . . . . . . . . . . 316
3.2.4.4 Pore Pressures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
3.2.5 Field Equations for Thermo- and Poro Elasticity . . . . . . . . . . . . . 317
3.3 Scalar Field: Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
3.3.1 Heat Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
3.3.2 Derivation of the Diusion Problem . . . . . . . . . . . . . . . . . . . . . 321
3.3.2.1 Simple 2D Derivation . . . . . . . . . . . . . . . . . . . . . . . . 321
3.3.2.2 Generalized Derivation . . . . . . . . . . . . . . . . . . . . . . . 322
3.3.2.3 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . 324
3.4 Summary and Tonti Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
4 MESH GENERATION
03
41
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 Triangulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2.1 Voronoi Polygon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2.2 Delaunay Triangulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2.3 MATLAB Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3 Finite Element Mesh Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.1 Boundary Denition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.2 Interior Node Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3.3 Final Triangularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
9.6.2.1 stiff.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
05
. 930
9.6.2.2 dmat.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 932
9.6.2.3 sfr.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 933
9.6.2.4 jacob.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 935
9.6.2.5 bmatps.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 935
9.6.3 Plott of Shape Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 936
. . . . . . 134
13.2.3.2 Reduced Integration . . . . . . . . . . . . . . . . . . . . . . . . . 135
13.2.3.3 Selective Reduced Integration . . . . . . . . . . . . . . . . . . . 136
13.3 Parasitic Shear/Incompatible Elements . . . . . . . . . . . . . . . . . . . . . . . . 137
13.3.1 Q4, The Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
13.3.2 Q6, The Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
13.3.3 QM6, Further Enhancements . . . . . . . . . . . . . . . . . . . . . . . . . 138
13.4 Rotational D.O.F. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
13.5 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1310
15 PLATES 151
15.1 Fundamental Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
15.1.1 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
15.1.2 Kinematic Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
15.1.3 Constitutive Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
15.2 Plate Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
15.2.1 Reissner-Mindlin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
15.2.1.1 Fundamental Relations . . . . . . . . . . . . . . . . . . . . . . . 156
15.2.1.2 Dierential Equation . . . . . . . . . . . . . . . . . . . . . . . . 158
15.2.1.3 Variational Formulation . . . . . . . . . . . . . . . . . . . . . . 158
15.2.2 Kirchho . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
15.2.2.1 Fundamental Relations . . . . . . . . . . . . . . . . . . . . . . . 159
15.2.2.2 Dierential Equation . . . . . . . . . . . . . . . . . . . . . . . . 1510
15.2.2.3 Stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1510
15.2.2.4 Variational Formulation . . . . . . . . . . . . . . . . . . . . . . . 1511
15.2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1511
15.3 Finite Element Formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1513
15.3.1 Rectangular Element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1513
15.3.1.1 Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1513
15.3.1.2 Shear Locking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1516
15.3.2 Nonconforming Kirchho Triangular Element . . . . . . . . . . . . . . . . 1517
15.3.2.1 Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
07
. 1517
15.3.2.2 Nonconformity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1519
15.3.3 Discrete Kirchho Triangle . . . . . . . . . . . . . . . . . . . . . . . . . . 1520
15.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1524
A VECTOR OPERATIONS A1
A.1 Vector Dierentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A1
A.1.1 Derivative WRT to a Scalar . . . . . . . . . . . . . . . . . . . . . . . . . . A1
E A-1 Tangent to a Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A2
A.1.2 Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A3
E A-2 Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A5
A.1.3 Gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A5
A.1.4 Scalar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A5
E A-3 Gradient of a Scalar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A6
E A-4 Stress Vector normal to the Tangent of a Cylinder . . . . . . . . . . . . . A6
A.2 Vector Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A9
A.2.1 Integral of a Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A9
A.2.2 Line Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A9
A.2.3 Integration by Parts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A9
A.2.4 Gauss; Divergence Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . A10
A.2.5 Stokes Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A10
A.2.6 Green; Gradient Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . A10
E A-5 Physical Interpretation of the Divergence Theorem . . . . . . . . . . . . A10
C MISC. C1
C.1 Units & Conversion Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C1
C.2 Metric Prexes and Multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . C2
List of Figures
5.1 Tonti Diagram for the Total Potential Energy, (Cervenka, J. 1994) . . . . . . . . 54
5.2 Tonti Diagram for Hu-Washizu, (Cervenka, J. 1994) . . . . . . . . . . . . . . . . 512
Draft
02
5.3 Uniformly Loaded Simply Supported Beam Analysed by the Rayleigh-Ritz Method515
LIST OF FIGURES
. 136
13.4 Hourglas Modes in Under-Integrated Quadratic Element . . . . . . . . . . . . . . 136
13.5 Rectangular Bilinear Element Subjected to Bending; Bilinear Element and Cor-
rect Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
13.6 Displacements Associated with Incompatible Modes for the Q6 Element . . . . . 138
13.7 Side Displacements Induced by Drilling d.o.f. . . . . . . . . . . . . . . . . . . . . 139
List of Tables
13.1 Full and Reduced Numerical Integrations for Quadrilateral Elements . . . . . . . 135
13.2 Bilinsear and Exact Displacements/Strains . . . . . . . . . . . . . . . . . . . . . 137
NOTATION
03
SCALARS
A Area
c Specic heat
E Elastic Modulus
h Film coecient for convection heat transfer
I Moment of inertia
J St Venants torsional constant
L Length
Q Rate of internal heat generation per unit volume
t Time
T Temperature
u, v, w Translational displacements along the x, y, and z directions
U0 Strain energy density
U Strain energy
U0 Complementary strain energy density
U Complementary strain energy
W Work
Potential energy
Coecient of thermal expansion
Shear modulus
Poissons ratio
mass density
Rotational displacement
M Virtual moment
P Virtual force
Virtual rotation
u Virtual displacement
Virtual curvature
U Virtual internal strain energy
W Virtual external work
TENSORS order 1
t Specied tractions along t
LIST OF TABLES
u Displacement vector
(x)
u Specied displacements along u
u Displacement vector
ue Nodal element displacements
u Nodal displacements in a continuous system
u Structure nodal displacements
V Shear forces in a plate Vx , Vy
Element nodal displacements
Virtual strain vector
Virtual stress vector
Stress vector
0 Initial stress vector
TENSORS order 2
TENSORS order 4
D Constitutive matrix
Surface
t Boundary along which surface tractions, t are specied
u Boundary along which displacements, u are specied
T Boundary along which temperatures, T are specied
c Boundary along which convection ux, qc are specied
q Boundary along which ux, qn are specied
Volume of body
FUNCTIONS, OPERATORS
Variational operator
B Discrete strain-displacement operator
L Linear dierential operator relating displacement to strains
Divergence, (gradient operator) on scalar
x
y
T
z
u
.u = div .u ux y
Divergence, (gradient operator) on vector x + y + u z
z
v 2 Euclidian norm.
. . . Innity norm.
PROGRAM ARRAYS
18 Mathematical Formulations.
2 23 Elasticity
25 Direct Method; Field Eq.
3 30 Variational Methods
Feb. 1 Mesh Generators; Laboratory
4 6 Variational Methods, Mechanics; Laboratory
8 FE Discretization and Requirements; C0 Elements.
5 13 Isoparametric Elements, Bar Element 6.1-6.14
15 Isoparamteric Element, Bilinear Element
6 20 Isoparameteric Element; Quadratic, Hierarchical Elements;
Numerical integration
22 Isoparameteric Element; Numerical integration
7 27 Laboratory
29 Weighted Residuals
8 Mar. 5 Galerkin; 3D Elasticity; Field Equation
7 Field Equation, Theory, application Ch. 16
9 12 Field Equation
14 Exam I
10 19 Lab (Field Equations)
21 Error Analysis
SPRING BREAK
12 Apr. 2 Topics (Condensation, Transformation, Integration, Test)
4 Order of Integration, Eignevalu tests
13 9 Plate Bending 9.1-9.6
11 Plate Bending 11.1-11.5
14 16 Plate Bending
18 Geometric Non Linearity Ch. 14.
15 23 Geometric Nonlinearity
25 Material Nonlinearity Ch. 17
16 30 Dynamics Ch. 13
May 2 Review
Chapter 1
PREREQUISITE
1 In the rst course (CVEN4525/5525, Finite Element I; Framed Structures), the direct stiness
method was rst introduced (element stiness matrix, transformation matrix, global stiness
matrix assembly, internal force recovery). As an interlude we then covered the exibility method
and stiness-exibility relationship. The second part of the course began with a thorough cov-
erage of variational method (duality between extremization of a functional and a corresponding
euler dierential equation) followed by a rigorous introduction/derivation of the various energy
methods.
3 The displacement at any point inside an element can be written in terms of the shape
functions N and the nodal displacements {}
= N{} (1.1)
Natural B.C.
Essential B.C.
div + b = 0 Du = 0 ij (ui,j + uj,i ) = 0
1
2
ij,j = 0
t t = 0 t u = 0 u ui u = 0 u ti = 0 t
def def
U0 = 0 d U0 = 0 d
Gauss Gauss
Principle of Virtual Work Principle of Complementary
Virtual Work
T d uT bd t uT
td = 0 ij ij d u ui ti d = 0
Wi We = 0
Wi We = 0
Principle of Stationary Principle of Complementary
Potential Energy Stationary Potential Energy
= 0 = 0
def def
= U We = Wi + We
= U0 d ( ui bi d + t ui ti d)
= U0 d + u ui ti d
Castiglianos Castiglianos
First Theorem Second Theorem
Wi k Wi k
=P =
k Pk
Rayleigh-Ritz
n
uj cji ji + j0
i=1
=0 i = 1, 2, , n; j = 1, 2, 3
cji
Principle of Stationary Principle of Virtual Work
Complementary Energy
Principle of Complementary Principle of Stationary
Virtual Work Potential Energy
Statically Admissible Stresses
Stresses satisfy the equilibrium conditions
and the static boundary conditions
Table 1.1: Summary of Variational Terms Associated with One Dimensional Elements
1 1
du L L
u1
= x = = (1.3-b)
dx u2
N1 N2
x x
{}
[B]
y d2 v
= =y 2 (1.4-a)
dx
1 M
= (1.4-b)
EI
d2 v
= y 2 (1.4-c)
dx
6 2 6 2 v1
L2 (2 1) L (3 2) L2 (2 + 1) L (3 1) 1
= y (1.4-d)
v2
2 N1 2 N2 2 N3 2 N4
x2 x2
x2 x2
2
[B]
{}
= [N]{} (1.5-a)
= [B]{} (1.5-b)
(1.5-c)
thus:
x = Ex Eix (1.7)
or in matrix form:
{} = [D]{} [D]{i } (1.8)
where [D] is the constitutive matrix which relates stress and strain.
U = W (1.9-a)
U = {}dvol (1.9-b)
vol
{} = [D]{} [D]{i } (1.9-c)
{} = [B]{} (1.9-d)
{} = [B]{} (1.9-e)
= [B] T
(1.9-f)
(1.9-g)
Combining Eqns. 1.9-a, 1.9-b, 1.9-c, 1.9-f, and 1.9-e, the internal virtual strain energy is
given by:
U = [B] [D][B]{} dvol
T
[B]T [D]{i }dvol
vol vol
= [B]T [D][B] dvol{} [B]T [D]{i }dvol (1.10-a)
vol vol
the virtual external work in turn is given by:
W = {F} + q(x)dx (1.11)
l
Virt. Nodal Displ. Nodal Force
Victor Saouma Finite Elements II; Solid Mechanics
Draft
16
{} = [N]{} (1.12)
yields:
l
W = {F} + [N]T q(x) dx (1.13)
0
Equating the internal strain energy Eqn. 1.10-a with the external work Eqn. 1.13, we obtain:
[B] [D][B] dvol{}
T
[B]T [D]{i }dvol =
vol vol
[K] {Finit }
l
{F} + [N]T q(x) dx (1.14-a)
0
{Fe }
where:
4 The physical interpretation of the global stiness matrix K is analogous to the one of the
element, i.e. If all degrees of freedom are restrained, then Kij corresponds to the force along
global degree of freedom i due to a unit positive displacement (or rotation) along global degree
of freedom j.
5 For instance, with reference to Fig. 1.3, we have three global degrees of freedom, 1 , 2 , and
P
17
2 1
0 3 P
1A
0
M
11
00 1
0 11A
00
0
1
P/2
00
11
A
0
1
111 11
000 00
EI
B B 0
1 00
11 B
1
w H
w
L/211
00 111
000
C C
0000
1111
L/2 0000
111100
11 000
111 C
11
00
K22 K32 0
1
K23
1
0
K 33
0
1 0
1
1 111K121
000 0
0
1 1111
0000
0 K13
1
1
0 A K211
0 K31 11
00A 01
1 1
0A B 1
0
1 B 1
0 00
11 B
0
1 0
1
0
1
1111
0000 0
1
0 K11
1 0
1
11
00
C 111
000 111
000
C
00
11 000
111
C
000
111
and the rst column corresponds to all the internal forces in the unrestrained d.o.f. when a
unit displacement along global d.o.f. 1 is applied.
6 The structural stiness matrix is assembled only for those active degrees of freedom which
are active (i.e unrestrained). It is the one which will be inverted (or rather decomposed) to
determine the nodal displacements.
7 The augmented stiness matrix is expressed in terms of all the dof. However, it is parti-
tioned into two groups with respective subscript u where the displacements are known (zero
otherwise), and t where the loads are known.
Pt Ktt Ktu t ?
= (1.20)
Ru ? Kut Kuu u
t = K1
tt (Pt Ktu u ) (1.21)
10 For internal book-keeping purpose, since we are assembling the augmented stiness matrix,
11 The element internal forces (axial and shear forces, and moment at each end of the member)
are determined from
(e)
pint = k(e) (e) (1.23)
(e)
at the element level where pint is the six by six array of internal forces, k(e) the element
stiness matrix in local coordinate systems, and (e) is the vector of nodal displacements in local
coordinate system. Note that this last array is obtained by rst identifying the displacements
in global coordinate system, and then premultiplying it by the transformation matrix to obtain
the displacements in local coordinate system.
1.3.2 Logistics
1.3.2.1 Boundary Conditions, [ID] Matrix
12 Because of the boundary condition restraints, the total structure number of active degrees of
freedom (i.e unconstrained) will be less than the number of nodes times the number of degrees
of freedom per node.
13 To obtain the global degree of freedom for a given node, we need to dene an [ID] matrix
such that:
ID has dimensions l k where l is the number of degree of freedom per node, and k is the
number of nodes).
ID matrix is initialized to zero.
1. At input stage read ID(idof,inod) of each degree of freedom for every node such that:
0 if unrestrained d.o.f.
ID(idof, inod) = (1.24)
1 if restrained d.o.f.
2. After all the node boundary conditions have been read, assign incrementally equation
numbers
Note that the total number of dof will be equal to the number of nodes times the number
of dof/node NEQA.
3. The largest positive global degree of freedom number will be equal to NEQ (Number Of
Equations), which is the size of the square matrix which will have to be decomposed.
1.3.2.2 LM Vector
PREREQUISITE
15 The LM vector of a given element gives the global degree of freedom of each one of the element
degree of freedoms. For the structure shown in Fig. 1.4, we would have:
LM = 1 2 4 5 6 7 element 1 (2 3)
LM = 5 6 7 1 2 3 element 2 (3 1)
LM = 1 2 3 3 8 9 element 3 (1 4)
16 As for the element stiness matrix, the global stiness matrix [K] is such that Kij is the
2. Share a node.
3. Are connected by an element and the corresponding value in the element stiness matrix
in the global coordinate system is zero.
20 There are usually more than one element connected to a dof. Hence, individual element
stiness matrices terms must be added up.
21 Because each term of all the element stiness matrices must nd its position inside the global
stiness matrix [K], it is found computationally most eective to initialize the global stiness
matrix [KS ](N EQAN EQA ) to zero, and then loop through all the elements, and then through
(e)
each entry of the respective element stiness matrix Kij .
(e)
22 The assignment of the element stiness matrix term Kij (note that e, i, and j are all known
since we are looping on e from 1 to the number of elements, and then looping on the rows and
S is made through
columns of the element stiness matrix i, j) into the global stiness matrix Kkl
the LM vector (note that it is k and l which must be determined).
23 Since the global stiness matrix is also symmetric, we would need to only assemble one side
of it, usually the upper one.
24 Contrarily to the previous method, we will assemble the full augmented stiness matrix.
50kN
4 kN/m
111
00
11
11
00
00
11
00
11
2 2 3
8m
3m
1
000
111
111
000
000
111
000
111
0001
111 7.416 m 8m
Finally the LM vectors for the two elements (assuming that Element 1 is dened from node 1
to node 2, and element 2 from node 2 to node 3):
4 5 6 1 2 3
[LM ] = (1.29)
1 2 3 7 8 9
Let us simplify the operation by designating the element stiness matrices in global coordinates
as follows:
4 5 6 1 2 3
4 A11 A12 A13 A14 A15 A16
5
A21 A22 A23 A24 A25 A26
6
A31 A32 A33 A34 A35 A36
K (1) = (1.30-a)
1
A41 A42 A43 A44 A45 A46
2 A51 A52 A53 A54 A55 A56
3 A61 A62 A63 A64 A65 A66
1 2 3 7 8 9
1 B11 B12 B13 B14 B15 B16
2
B21 B22 B23 B24 B25 B26
3
B31 B32 B33 B34 B35 B36
K (2) = (1.30-b)
7
B41 B42 B43 B44 B45 B46
8 B51 B52 B53 B54 B55 B56
9 B61 B62 B63 B64 B65 B66
We note that for each element we have shown the corresponding LM vector.
A44 + B11 A45 + B12 A46 + B13 A41 A42 A43 B14 B15 B16
A54 + B21 A55 + B22 A56 + B23 A51 A52 A53 B24 B25 B26
A64 + B31 A65 + B32 A66 + B33 A61 A62 A63 B34 B35 B36
A14 A15 A16 A11 A12 A13 0 0 0
K=
A24 A25 A26 A21 A22 A23 0 0 0
(1.31)
A34 A35 A36 A31 A32 A33 0 0 0
B41 B42 B43 0 0 0 B44 B45 B46
B51 B52 B53 0 0 0 B54 B55 B56
B61 B62 B63 0 0 0 B64 B65 B66
We note that some terms are equal to zero because we do not have a connection between
the corresponding degrees of freedom (i.e. node 1 is not connected to node 3).
1.3.2.4 Algorithm
1. Identify type of structure (beam, truss, grid or frame) and determine the
(a) Number of spatial coordinates (1D, 2D, or 3D)
(b) Number of degree of freedom per node (local and global)
(c) Number of cross-sectional and material properties
2. Determine the global unrestrained and restrained degree of freedom equation num-
bers for each node, Update the [ID] matrix (which included only 0s and 1s in the
input data le).
Analysis :
7. For each element, transform its nodal displacement from global to local coordinates
113
26 Some of the prescribed steps are further discussed in the next sections.
Using the direct stiness method, analyze the truss shown in Fig. 1.6.
4 4 5 50k
5
6
1 8
3 12
1 2 2 7 3
100k
16 16
Figure 1.6:
Solution:
0 0 1 0 0
ID = (1.32-a)
1 0 1 0 0
N ode 1 2 3 4 5
1 2 2 4 6
= (1.32-b)
1 3 3 5 7
LM 1 = 1 1 4 5 (1.33-a)
LM 2 = 1 1 2 3 (1.33-b)
LM 3 = 2 3 4 5 (1.33-c)
LM 4 = 4 5 6 7
PREREQUISITE
(1.33-d)
LM 5 = 2 3 4 5 (1.33-e)
LM 6 = 2 3 6 7 (1.33-f)
LM 7 = 2 3 2 3 (1.33-g)
LM 8 = 2 3 6 7 (1.33-h)
3. Determine the element stiness matrix of each element in the global coordinate system noting
that for a 2D truss element we have
1 1 2 3
1 18, 750 0 18, 750 0
1
0 0 0 0
[K2 ] = (1.36)
2 18, 750 0 18, 750 0
3 0 0 0 0
EA
Element 3 L = 12 , c = 0 , s = 1 , L = 25, 000 k/ft.
2 3 4 5
2 0 0 0 0
3
0 25, 000 0 25, 000
[K3 ] = (1.37)
4 0 0 0 0
5 0 25, 000 0 25, 000
EA
Element 4 L = 16 , c = 1 , s = 0 , L = 18, 750 k/ft.
4 5 6 7
4 18, 750 0 18, 750 0
5
0 0 0 0
[K4 ] = (1.38)
6 18, 750 0 18, 750 0
7 0 0 0 0
Element 5 L = 20 , c =
160
= 0.8 , s = 0.6 , EA
= 15, 000 k/ft.
115
20 L
2 3 4 5
2 9, 600 7, 200 9, 600 7, 200
3
7, 200 5, 400 7, 200 5, 400
[K5 ] = (1.39)
4 9, 600 7, 200 9, 600 7, 200
5 7, 200 5, 400 7, 200 5, 400
EA
Element 6 L = 20 , c = 0.8 , s = 0.6 , L = 15, 000 k/ft.
2 3 6 7
2 9, 600 7, 200 9, 600 7, 200
3
7, 200 5, 400 7, 200 5, 400
[K6 ] = (1.40)
6 9, 600 7, 200 9, 600 7, 200
7 7, 200 5, 400 7, 200 5, 400
EA
Element 7 L = 16 , c = 1 , s = 0 , L = 18, 750 k/ft.
2 3 2 3
2 18, 750 0 18, 750 0
3
0 0 0 0
[K7 ] = (1.41)
2 18, 750 0 18, 750 0
3 0 0 0 0
EA
Element 8 L = 12 , c = 0 , s = 1 , L = 25, 000 k/ft.
2 3 6 7
2 0 0 0 0
3
0 25, 000 0 25, 000
[K8 ] = (1.42)
6 0 0 0 0
7 0 25, 000 0 25, 000
4. Assemble the global stiness matrix in k/ft Note that we are not assembling the augmented
stiness matrix, but rather its submatrix [Ktt ].
8 9 2 38 9
>
>
>
0 >
>
>
9, 600 + 18, 750 18, 750 0 9, 600 7, 200 0 0 >
>
>
u1 >
>
>
>
> 0 >
> 6 9, 600 7, 200 7 >
> u2 >
>
= 6
9, 600 + (2) 18, 750 7, 200 0 0
>
>
< 100k
>
> 6 5, 400 + 25, 000 0 25, 000 7, 200 5, 400
7
7 >
>
< v3
>
>
=
0
6
=6 18, 750 + (2)9, 600 7, 200 7, 200 18, 750 0
7
7 u4
>
>
> >
>
> 6
6 7
7 >
> >
>
5> >
> 0 > SYMMETRIC 25, 000 + 5, 400(2) 0 0 > v5 >
>
>
> >
>
> 4 >
>
> u6 >
>
>
: 50k ; 18, 750 + 9, 600 7, 200 : ;
0 25, 000 + 5, 400 v7
(1.43)
0
2, 362.5 1, 562.5 0 800 600 0 0
U1
800 600
0 3, 925.0 600 0 0 U2
100 2, 533.33 0 2, 083.33 600 450
V3
0 = 3, 162.5 0 1, 562.5 0
U4
SYMMETRIC
0
2, 983.33 0 0 V5
50
2, 362.5 600
U6
0 2, 533.33 V7
Pt Ktt ut
(1.44)
U1
0.0223 in
U 2
0.00433 in
V3 0.116 in
U4 = 0.0102 in (1.45)
V5
0.0856 in
U6
0.00919 in
V7 0.0174 in
7. Solve for member internal forces (in this case axial forces) in local coordinate systems
U 1
u1 c s c s V1
= (1.46)
u2 c s c s U
2
V2
Element 1
1
0.0223
p1 1 ft 0.8 0.6 0.8 0.6 0
= (15, 000 k/ft)( ) (1.47-a)
p2 12 in 0.8 0.6 0.8 0.6 0.0102
0.0856
52.1 k
= Compression (1.47-b)
52.1 k
Element 2
2
0.0233
p1 1 ft 1 0 1 0 0
= 18, 750 k/ft( ) (1.48-a)
p2 12 in 1 0 1 0 0.00433
0.116
43.2 k
= Tension (1.48-b)
43.2 k
Element 3
3
0.00433
p1 1 ft 0 1 0 1 0.116
= 25, 000 k/ft( ) (1.49-a)
p2 12 in 0 1 0 1 0.0102
0.0856
63.3 k
= Tension (1.49-b)
63.3 k
Element 4
117
4
0.0102
p1 1 ft 1 0 1 0 0.0856
= 18, 750 k/ft( ) (1.50-a)
p2 12 in 1 0 1 0 0.00919
0.0174
1.58 k
= Tension (1.50-b)
1.58 k
Element 5
5
0.0102
p1 1 ft 0.8 0.6 0.8 0.6 0.0856
= 15, 000 k/ft( ) (1.51-a)
p2 12 in 0.8 0.6 0.8 0.6 0
0
54.0 k
= Compression (1.51-b)
54.0 k
Element 6
6
0.00433
p1 1 ft 0.8 0.6 0.8 0.6 0.116
= 15, 000 k/ft( ) (1.52-a)
p2 12 in 0.8 0.6 0.8 0.6 0.00919
0.0174
60.43 k
= Tension (1.52-b)
60.43 k
Element 7
7
0.00433
p1 1 ft 1 0 1 0 0.116
= 18, 750 k/ft( ) (1.53-a)
p2 12 in 1 0 1 0 0
0
6.72 k
= Compression (1.53-b)
6.72 k
Element 8
8
0
p1 1 ft 0 1 0 1 0
= 25, 000 k/ft( ) (1.54-a)
p2 12 in 0 1 0 1
0.00919
0.0174
36.3 k
= Compression (1.54-b)
36.3 k
Victor Saouma Finite Elements II; Solid Mechanics
Draft
118 PREREQUISITE
The simple frame shown in Fig. 1.7 is to be analyzed by the direct stiness method. Assume:
E = 200, 000 MPa, A = 6, 000 mm2 , and I = 200 106 mm4 . The complete MATLAB solution
is shown below along with the results.
50kN
4 kN/m
0
1
1
0
0
1
0
1
2 2 0
1
3
8m
3m
1
111
000
000
111
000
111
0001
111
000
111 7.416 m 8m
-4 -5 -6 1 2 3;
119
1 2 3 -7 -8 -9];
% Assemble augmented stiffness matrix
Kaug=zeros(9); for elem=1:2
for r=1:6
lr=abs(LM(elem,r));
for c=1:6
lc=abs(LM(elem,c));
Kaug(lr,lc)=Kaug(lr,lc)+K(r,c,elem);
end
end
end
% Extract the structures Stiffness Matrix
Ktt=Kaug(1:3,1:3);
% Determine the fixed end actions in local coordinate system
fea(1:6,1)=0; fea(1:6,2)=[0,8*4/2,4*8^2/12,0,8*4/2,-4*8^2/12];
% Determine the fixed end actions in global coordinate system
FEA(1:6,1)=Gamma(:,:,1)*fea(1:6,1); FEA(1:6,2)=Gamma(:,:,2)*fea(1:6,2);
% FEA_Rest for all the restrained nodes
FEA_Rest=[0,0,0,FEA(4:6,2)];
% Assemble the load vector for the unrestrained node
P(1)=50*3/8;P(2)=-50*7.416/8-FEA(2,2);P(3)=-FEA(3,2);
% Solve for the Displacements in meters and radians
Displacements=inv(Ktt)*P
% Extract Kut
Kut=Kaug(4:9,1:3);
% Compute the Reactions and do not forget to add fixed end actions
Reactions=Kut*Displacements+FEA_Rest
% Solve for the internal forces and do not forget to include the fixed end actions
dis_global(:,:,1)=[0,0,0,Displacements(1:3)];
dis_global(:,:,2)=[Displacements(1:3),0,0,0]; for elem=1:2
dis_local=Gamma(:,:,elem)*dis_global(:,:,elem);
int_forces=k(:,:,elem)*dis_local+fea(1:6,elem)
end
function [k,K,Gamma]=stiff(EE,II,A,i,j)
% Determine the length
L=sqrt((j(2)-i(2))^2+(j(1)-i(1))^2);
% Compute the angle theta (careful with vertical members!)
if(j(1)-i(1))~=0
alpha=atan((j(2)-i(2))/(j(1)-i(1)));
else
alpha=-pi/2;
end
% form rotation matrix Gamma
Gamma=[ cos(alpha) sin(alpha) 0 0 0 0; -sin(alpha)
cos(alpha) 0 0 0 0; 0 0 1 0 0
0; 0 0 0 cos(alpha) sin(alpha) 0; 0 0
0 -sin(alpha) cos(alpha) 0; 0 0 0 0
PREREQUISITE
0
1];
% form element stiffness matrix in local coordinate system
EI=EE*II; EA=EE*A; k=[EA/L, 0, 0, -EA/L, 0, 0;
0, 12*EI/L^3, 6*EI/L^2, 0, -12*EI/L^3, 6*EI/L^2;
0, 6*EI/L^2, 4*EI/L, 0, -6*EI/L^2, 2*EI/L;
-EA/L, 0, 0, EA/L, 0, 0;
0, -12*EI/L^3, -6*EI/L^2, 0, 12*EI/L^3, -6*EI/L^2;
0, 6*EI/L^2, 2*EI/L, 0, -6*EI/L^2, 4*EI/L];
% Element stiffness matrix in global coordinate system
K=Gamma*k*Gamma;
Displacements =
0.0010
-0.0050
-0.0005
Reactions =
130.4973
55.6766
13.3742
-149.2473
22.6734
-45.3557
int_forces = int_forces =
141.8530 149.2473
2.6758 9.3266
13.3742 -8.0315
-141.8530 -149.2473
-2.6758 22.6734
8.0315 -45.3557
We note that the internal forces are consistent with the reactions (specially for the second node
of element 2), and amongst themselves, i.e. the moment at node 2 is the same for both elements
(8.0315).
v1 1 v2 2
V1 12EI/L3 6EI/L2 12EI/L3 6EI/L2
M1 6EI/L2 4EI/L 6EI/L2 2EI/L
[ke ] = (1.56)
V2 12EI/L3 6EI/L2 12EI/L3 6EI/L2
M2 6EI/L2 2EI/L 6EI/L2 4EI/L
This matrix is singular, it has a rank 2 and order 4 (as it embodies also 2 rigid body motions).
27 We shall consider 3 dierent cases, Fig. 1.8
1
0 1
0 1
0
P
-3 -3
0
1 0
1 0
1
-4 1 1 -4
2 2
1
0 0
1 0
1 0
1
M
0
1
0
1
0
1
-2 -3
1
0
0
1
-4
1
0
1
0
1 0
1
0
1
0
1
4. Ktt is inverted (or actually decomposed) and stored in the same global matrix
2 3
L3 /3EI L2 /2EI 12EI/L3 6EI/L2
6 7
6 7
6 L2 /2EI L/EI 6EI/L2 2EI/L 7
6 7
4 12EI/L3 6EI/L2 12EI/L3 6EI/L2 5
6EI/L2 2EI/L 6EI/L2 4EI/L
8 9 2 3
8 9
>
> P >
> L3 /3EI L2 /2EI 12EI/L3 6EI/L2 P >
>
< >
= 6 7>>
< >
6 7 0 =
Pt Ktu u = 0 6
6
L2 /2EI L/EI 6EI/L2 2EI/L 7
7> 0 >
>
> >
0 > 5> >
> ; 4 12EI/L2
> : ;
3 2 3
: 6EI/L 12EI/L 6EI/L2
0 6EI/L 2EI/L 6EI/L2 4EI/L 0
8 9
>
> P >
>
>
< >
=
= 0
>
> 0 >
>
>
: >
;
0
>
> >
>
>
> 0 >
>
: ;
0
4. Ktt is inverted
123
2 3
L3 /3EI L/6EI 6EI/L2 6EI/L2
6 7
6 L/6EI 7
6 L/3EI 6EI/L2 6EI/L2 7
6 7
4 6EI/L2 6EI/L2 12EI/L3 12EI/L3 5
6EI/L 2
6EI/L2 12EI/L3 12EI/L3
5. We compute the equivalent load, P t = Pt Ktu u , and overwrite Pt by P t
2 3
8 9 L3 /3EI L/6EI 6EI/L2 6EI/L2
8 9
>
> 0 >
> 6 7>>
0 >
>
< = 6 7 < M =
Pt Ktu u = M 6
6
L/6EI L/3EI 6EI/L2 6EI/L2 7
7> 0 >
>
> >
> 4 5> >
: 0 ; 6EI/L2 6EI/L2 12EI/L3 12EI/L3 : ;
0 6EI/L2 6EI/L2 12EI/L3 12EI/L3 0
8 9
>
> 0 >
>
< =
= M
>
> >
>
: 0 ;
0
8 9 2 38 9
>
> M >
> 4EI/L 6EI/L2 2EI/L 6EI/L2 >
> 1 ? >
>
< = 6 6EI/L2 < =
12EI/L3 6EI/L2 12EI/L3 7
=6 7
R2 ? 2
>
> R ? > 4 2EI/L 6EI/L2 6EI/L 2 5
> >
: 3 > > >
4EI/L
; : 3 ;
R4 ? 6EI/L2 12EI/L3 6EI/L2 12EI/L3 4
4. Ktt is inverted (or actually decomposed) and stored in the same global matrix
2 3
L/4EI 6EI/L2 2EI/L 6EI/L2
6 7
6 6EI/L2 12EI/L3 6EI/L2 12EI/L3 7
6 7
4 2EI/L 6EI/L2 4EI/L 6EI/L2 5
6EI/L2 12EI/L3 6EI/L2 12EI/L3
Chapter 2
INTRODUCTION
2.1 Introduction
1 Whereas the rst course focused exclusively on one dimensional rod elements, this course
will greatly expand our horizons by considering introducing a methodology to solve partial
dierential equations, with special emphasis on solid mechanics.
2 The eld of mechanics, can itself be subdivided into four major disciplines:
Theoretical which deals with the fundamental laws and principles of mechanics. A Continuum
Mechanics course is a must.
Applied mechanics seeks to apply the theoretical knowledge to engineering applications. Elas-
ticity or Fracture Mechanics solutions are such an example of applied mechanics.
Computational mechanics combines mathematical models with numerical methods to solve
problems on a digital computer.
Experimental mechanics is conducted exclusively in a laboratory through physical measure-
ments.
3 Any problem characterized by a PDE can be analyzed by the nite element method. The
4 Since the nite element method is a numerical scheme to solve (partial) dierential equations,
let us closely examine some of the major PDE which can be solved.
5The general form of a partial dierential equation is (note that we adopt the tensor notation
where u,x = du ):
dx
F (x, y, z, , u, u,x , u,y , u,z , , u,xx , u,yy , , u,xy , u,xz , ) = 0
and the order of the PDE is dened by the order of the highest partial derivatives appearing
in the equation. For instance
1 u,xx + 2 u,xy + 3 u,yy + 4 = 0
Draft
22 INTRODUCTION
Change
Physical Problem Physical Problem
Mathematical Model
Governed by Differential Equations
Assumptions on
Geometry
Kinematics Improve
Material Mathematical Model
Loading
Boundary Conditions
Etc.
FEM Solution of Mathematical Model
Design Improvements
Structural Optimization
6 Many signicant physical systems can be described by second order linear partial dierential
equations. The most general form is
2u 2u 2u u u
A(x, y) 2 + 2B(x, y) + C(x, y) 2 = x, y, u, , (2.1)
x xy y x y
8 Note:
1. The Laplace equation (2 u = 0) is a special case of Poissons equation, where the right
hand side is zero. Laplace associated with the equilibrium problem.
9 This classication is established when solving Eq. 2.1 using the method of characteristics
because it is then observed that the character of the solutions is distinctly dierent for the
three categories of equations.
The idealized dam shown in Fig. 2.2 stands on permeable soil. Formulate the dierential
111
000
000
111
000
111
000
111
000
111
000
111
Impermeable wall
000
111
h1 000
111
000
111 q(y+dy)
000
111
000
111
000
111 h2
q(x) q(x+dx)
dy
h
Seepage
L q(y)
y Permeable soil dx
x
0000000000000000000000000
1111111111111111111111111
1111111111111111111111111
0000000000000000000000000
Impermeable rock
equation governing the steady-state seepage of water through the soil and give the corresponding
boundary conditions.
Solution:
1. For a typical element of widths dx and dy (and unit thickness), the total ow into the
element must be equal to the total ow out of the element. Hence we have
qy qx
qy qy + dx + qx qx + dy = 0 (2.2-a)
y x
qy qx
dydx dxdy = 0 (2.2-b)
y x
2. Using Darcys law, the ow is given in terms of the total potential (water elevation),
qx = k qy = k (2.3)
x y
where we assume a uniform permeability k. Substituting from Eq. 2.3 into Eq. 2.2-b, we
obtain the Laplace equation
2 2
k + 2 =0 (2.4)
x2 y
3. It may be noted that this same equation is also obtained in heat transfer analysis and in
the solution of electrostatic potential and other eld problems.
4. The boundary conditions are no-ow boundary conditions in the soil at x = and
x = +,
= 0; =0 (2.5)
x x= x x=+
at the rock-soil interface,
=0 (2.6)
y y=0
and at the dam-soil interface,
=0 (2.7)
y h x+ h ,y=L
2 2
The dierential equation in Eq. 2.4 and the boundary conditions in Eq. 2.5 to Eq. 2.8
dene the seepage ow steady-state response.
y 11111111
00000000
25
111
000 00000000
11111111
000
111 00000000
11111111
000111
111000 111
000 00000000
11111111
00000000
11111111
000
111
000
111 000
111
000
111 00000000
11111111
1
0 00000000
11111111
00
11
0
1
0
1 00
11
z
0
1 00
11
0q(x)
1 00
11
00
11
dz=1.0
q (t) x 0
1
0
1 00
11 q(x+dx)
1.0
00
11
0
t>0 0
1
0
1 00
11
0
1 00
11
111
000 0
1 00
11
.0
0
1 00
11
=1
000
111 0
1 00
11
dy
000
111 00
11
L dx
The very long slab shown in Fig. 2.3 is at a constant initial temperature i when the surface
at x = 0 is suddenly subjected to a constant uniform heat ow input. The surface at x = L of
the slab is kept at the temperature i , and the surfaces parallel to the x, z plane are insulated.
Assuming one-dimensional heat ow conditions, show that the problem-governing dierential
equation is the heat conduction equation
2
k 2
= c (2.9)
x t
where is the mass density, c is the heat capacity per unit mass (amount of heat needed to raise
a unit mass by one degree), k is the conductivity, and the temperature is the state variable.
State also the boundary and initial conditions.
Solution:
1. We consider a typical dierential element of the slab. The element equilibrium require-
ment is that the net heat ow input to the element must equal the rate of heat stored in
the element. Thus
q
qA q + dx A = Ac (2.10)
x t
2
k 2
= c (2.12)
x t
1
0 F(t)
0
1
0
1 x u(x,t)
F0
0 11
1 00 00
11 11
00 00 11 00
11 F0
A A + d d x x+dx t
x dx dx
In this case the element interconnectivity requirements are contained in the assumption
that the temperature be a continuous function of x and no additional compatibility
conditions are applicable.
The rod shown in Fig. 2.4 is initially at rest when a load F (t) is suddenly applied at its
free end. Show that the problem-governing dierential equation is the wave equation
2u 1 2u E
2
= 2 2 ; and c = (2.15)
x c t
where E is the Youngs modulus, the mass density, and A the cross sectional area, c corre-
sponds to the velocity of sound in the elastic medium, and u is the state variable. Also state
the boundary and initial conditions.
Solution:
1. The element force equilibrium requirements of a typical dierential element are, using
dAlemberts principle 1 which states that with inertia forces included, a system is in
1
Thinking in terms of equilibrium of forces, it is more appealing to invoke DAlemberts principle of dy-
namic equilibrium rather than Newtons second law of motion. This principle is based on the notion of a
ctitious inertia force, equal to the product of mass times acceleration and acting in a direction opposite to
the acceleration.
2u
A + A dx A = A 2 dx (2.16)
x t
2u 1 2u
= (2.18)
x2 c2 t2
4. The element interconnectivity requirements are satised because we assume that the dis-
placement u is continuous, and no additional compatibility conditions are applicable.
u|(0,t) = 0 (2.19-a)
u
EA = F0 (2.19-b)
x (L,t)
(2.19-c)
u(x,0) = 0 (2.20-a)
u
= 0 (2.20-b)
t
(x,0)
With the above two sets of conditions, the formulation of the problem is complete, and
Eq. 2.18 can be solved for the displacement response of the rod.
10 Observations
1. The unknown state variables (or their normal derivatives) are given on the boundary.
These problems are for this reason also called boundary value problems, where
we should note that the solution at a general interior point depends on the data at
every point of the boundary.
2. A change in only one boundary value aects the complete solution; for instance, the
complete solution for depends on the actual value of h1 .
3. Elliptic dierential equations generally govern the steady-state response of systems.
1. Comparing the governing dierential equations given in the three examples we note
that in contrast to the elliptic equation, the parabolic and hyperbolic equations
Eq. 2.12 and Eq. 2.18 include time as an independent variable and thus dene
propagation problems. These problems are also called initial value problems
because the solution depends on the initial conditions.
2. We may note that analogous to the derivation of the dynamic equilibrium equations
of lumped-parameter models, the governing dierential equations of propagation
problems are obtained from the steady-state equations by including the resistance
to change (inertia) of the dierential elements.
3. The parabolic and hyperbolic dierential equations Eq. 2.12 and Eq. 2.18 would
become elliptic equations if the time-dependent terms were neglected. In this way
the initial value problems would be converted to boundary value problems with
steady-state solutions.
4. The solution of a boundary value problem depends on the data at all points of the
boundary. However, in propagation problem, the solution at an interior point may
depend only on the boundary conditions of part of the boundary and the initial
conditions over part of the interior domain.
4. Calculation of state variables: The set of linear equations is solved to determine the
state variables at each discretized points.
12 In the following sections, we shall illustrate, through a number of dierent physical problems,
13 In this rst class of problem, we shall focus on equilibrium problems, that is problems
14 We seek to determine the displacements of each of the rigid carts connected by linear elastic
springs, as well as the force in each spring, Fig. 2.5. Conceptually, this can be viewed as a one
dimensional truss, with the spring stiness k corresponding to AE/L
u 1, R 1 k4 u 2, R 2 u 3, R 3
1111111
0000000
0000000
1111111 1111111
0000000
0000000
1111111
0000000
1111111 k1111111
0000000
0000000
1111111 0000000
1111111
k1 0000000
1111111
0000000
1111111
3
1111111
0000000 k5 0000000
1111111
0000000
1111111
11
00 0000000
1111111 0000000
1111111 0000000
1111111
00
11 0000000
1111111 0000000
1111111 0000000
1111111
11
00 0000000
1111111 0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111 0000000
1111111
0000000
1111111
1
0000000
1111111
k2
0000000
1111111
2 0000000
1111111
3
0000000
1111111
0000000
1111111 0000000
1111111 0000000
1111111
0000000
1111111
0000000
1111111 0000000
1111111 0000000
1111111
0000000
1111111 0000000
1111111 0000000
1111111
Figure 2.5: System of Rigid Carts Interconnected by Linear Springs, (Bathe 1996)
1. Considering the rst two springs, we can write the following equations of equilibrium
(1)
k1 u1 = F1 (2.21-a)
(2)
k2 (u1 u2 ) = F1
(2) (2.21-b)
k2 (u2 u1 ) = F2
2. By extension, we can write the equation of equilibrium for each spring in terms of the
displacement state variables
(2)
F1 1 1 u1
= k2 (2.22-a)
F2
(2) 1 1 u2
(4)
F1 1 1 u1
= k4 (2.22-b)
F3
(4) 1 1 u3
(3)
F1 1 1 u1
= k3 (2.22-c)
F2
(3) 1 1 u2
(5)
F2 1 1 u2
= k5 (2.22-d)
F3
(5) 1 1 u3
4. We next substitute the equilibrium equations of each element into the three global equa-
tions of equilibrium
(k1 + k2 + k3 + k4 ) (k2 + k3 ) k4 u1 R1
K= (k2 + k3 ) (k2 + k3 + k5 ) k5 u = R (2.24)
2 2
k4 k5 (k4 + k5 ) u3 R3
5. The state variables (nodal displacements) ui can be solved through the inversion of K.
6. The forces in each spring can be determined from equilibrium of each individual spring.
temperatures in the wall is characterized by the external surface temperature 1 and 3 and the
interface temperature 2 . Establish the equilibrium equations of the problem in terms of these
temperatures when the ambient temperatures 0 and 4 are known.
0 1 2 3
4
111
000 000
111 000
111
000
111
000
111 000
111
000
111 000
111
000
111
0
1
000000000
111111111
111111111
0
1 0
1
000000000111111111111111111
000000000000000000
000000000
111111111
000000000
111111111
0
1
000000000
111111111
000000000
111111111
0
1
000000000000000000
111111111111111111
0
1
000000000000000000
111111111111111111
0
1
000000000
111111111
000000000
111111111
0
1 0
1
000000000000000000
111111111111111111
00
11
3k 00
11
000000000
111111111
000000000
111111111
0
1
000000000
111111111
000000000
111111111
0
1
000000000000000000
111111111111111111
0
1 00
11
000000000000000000
111111111111111111
2k
0
1
000000000
111111111
000000000
111111111 0
1
000000000000000000
111111111111111111
0
1
000000000
111111111
000000000
111111111 0
1
000000000000000000
111111111111111111
0
1
000000000
111111111
000000000
111111111
0
1 0
1
000000000000000000
111111111111111111
000000000
111111111
000000000
111111111
0
1
0
1
000000000000000000
111111111111111111
000000000
111111111
000000000
111111111
0
1
000000000
111111111
A 111111111
B C 0
1
000000000000000000
111111111111111111
0
1
000000000000000000
111111111111111111 D
000000000
0
1
000000000
111111111
000000000
111111111
0
1 0
1
000000000000000000
111111111111111111
000000000
111111111
000000000
111111111
0
1
000000000
111111111
2k
000000000
111111111
0
1
000000000000000000
111111111111111111
000000000000000000
111111111111111111
3k 0
1
0
1
000000000
111111111
000000000
111111111000000000000000000
111111111111111111
0
1
0
1
000000000
111111111000000000000000000
0
1
111111111111111111
Conductance
q = Ak (2.25)
where q is the total heat ow, A the area, the temperature drop in the direction of
heat ow, and k the conductance or surface coecient.
q1 = q2 = q3 = q4 (2.27)
thus
211
16 In this example, we seek to establish the equations that govern the steady-state pressure
and ow distribution in the hydraulic network shown in Fig. 2.7. We assume the uid to be
incompressible and the pressure drop in a branch to be proportional to the ow q through that
branch, (Darcys law)
d4 p
q= p = Rq (2.30)
128L
where d is the pipe diameter, the uid kinematic viscosity, L the pipe length, R is the branch
eective resistance coecient.
E
q R=10b
1
A B
Q R= R=2b q b Q
3 5
5b R=
q q
2 2
C R=3b q D
4
1. We consider each branch of the pipe to constitute an element, and we select as state
variables pA , pC , and pD the pressures at A, C, and D respectively.
Q = q1 + q2
q2 |AC = q3 + q4 (2.32)
q2 |DB = q3 + q4
Substituting, we get
3 2 0 pA 10bQ
6 31 25 p = 0 (2.33)
C
1 1 1 pD 0
or
9 6 0 pA 30bQ
6 31 25 p = 0 (2.34)
C
0 25 31 pD 0
4. The pressures are determined through the inversion of the matrix, and the ow can in
turn be determined from Eq. 2.31
2.3.1.4 DC Network
17 Considering the network shown in Fig. 2.8, determine the steady state current distribution
in the network.
I3 I3
6R
I3 4
I - I3 2R I - I3 4R
1 2
2 3
1 2R I2
I
1 B A
2E E
1. The state variables will be the currents I1 , I2 , and I3 . Ohms law will be applied
E = RI (2.35)
4R 0 2R I1 2E
0 4R 4R I = E (2.37)
2
2R 4R 12R I3 0
u1 , R1 u2 , R2 u3 , R3
01 1
0
1010 1 2 3 4 0
1
0
1
1010 0
1
3K 2K 3K 2K 0
1
1
0110 1
0
0
1
100 10 b
0
1
11 0
1
2 2b 3 5
1010 20
110 1
0
31
0 0
1
1010 1010 0
1 0
1
0
1
5b
3b 0
1 5b 0
1
4
u1 , R1 u2 , R2 u3 , R3
u2 , R2
1
0 3
1
0
21
0 0
1 4
0
1 0
1 1
0
0
1 4R 31
0 0
1
0
1
01
1
11
0 2 0
1 0
1
1010 0
1 0
1
0
1
6R
0
1 0
1
10 2R 0
1 2R
0
1
u1 , R1 u3 , R3
(1)
Spring 1: 3Ku1 = F
1
(1)
1 1 u1 F1
Spring 2: 2K =
1 1 u2 F2
(2)
(2.38)
(3)
1 1 u2 F2
Spring 3: 3K =
1 1 u3 F3
(3)
(4)
Spring 4: 2Ku3 = F3
= R1
(1) (2) (2) (1)
F1 + F1 F1 = R1 F1 = Q 10b
1
u1
(2) (3) (4)
F3 + F3 + F3 = R2 R2 = 0
(3) (4) (5)
(3) (4) (5)
F4 + F4 + F4 = R3 F4 + F4 = R3 F4 = 0 5b
1
u3
(2.44)
4. Rearrange state variables and multiply by 30
9 6 0 u1 30bQ
6 31 25 u = 0 (2.45)
2
0 25 31 u3 0
(4)
Spring 4: 6Ru3 = F3
18 The main characteristic of a propagation dynamic problem is that the response of the system
changes with time. In principle, we may apply the same analysis procedure as in steady-state
problems, however in this case the state variables and the the equilibrium relations depend on
time.
Let us reconsider the system of rigid carts previously analysed.We assume the springs to be
massless, but the carts have masses mi .
where
d2 ui
ui = for i = 1, 2, 3 (2.51)
dt2
2. Thus we obtain the following system of governing equilibrium equations
Mu + Ku = R(t) (2.52)
where
m1 0 0
M = 0 m2 0 (2.53)
0 0 m3
1. The state variables are the temperature of the gas, 1 , and the temperature of the wall
2 .
C + K = Q (2.55-a)
C1 0
C = (2.55-b)
0 C2
Victor Saouma Finite Elements II; Solid Mechanics
Draft
2.3 Solution of Discrete-System Mathematical models 217
Wall
1
0 1
0
0
1 0
1 Gas 1
0
1 0
1
Co
0
1 0
1
1
k
nv
Atmosphere
0
1 0
1
on
ec
0
1 0
1
c ti
tio
Gas
0
1 0
1
ve
n
0
1 0
1
k2
Co
0
1 0
1
0
1 0
1 Radiation kr Convection
0
1 0
1 Filament f Wall 2 Atmosphere a
0
1
0
1 0
1
0
1
k3
0
1 0
1
0
1 0
1
0
1 0
1
0
1 0
1
0
1 0
1
Filament
(k1 + k2 ) k2
K = (2.55-c)
k2 (k2 + k3 )
1
= (2.55-d)
2
! k1 f "
Q = (2.55-e)
kr (f )4 (2 )4 + k3 a )
4. We note that because of the radiation boundary conditions, the heat ow equilibrium
equations are nonlinear in .
Mu + Ku = 0 (2.56)
where the motion is referred to being free, since there are no applied loads.
20 By assuming a harmonic motion
u = sin t (2.57)
the natural frequencies and the corresponding mode shapes can be determined from
the generalized eigenvalue problem
2 M = K (2.58)
or
(K 2 M) = 0 (2.59)
Since is nontrivial
|K 2 M| = 0 (2.60)
or with 2 =
INTRODUCTION
|K M| = 0 (2.61)
which is the characteristic equation, and is called the eigenvalue of the equation, and the
structure is said to respond in the mode corresponding to a particular frequency.
21For computational purposes, if we premultiply each side of the preceding equation by M1 ,
then
(M1 K I) = 0 (2.62)
It should be noted that a zero eigenvalue is obtained for each possible rigid body motion of a
structure that is not completely supported.
22 Depending on which mass matrix is adopted, slightly dierent results are obtained. In
general, lumped mass matrices approach the exact value (consistent mass matrix) from below.
23 The mode shapes are shapes , anfd give a relative magnitude of the DOF, not the
absolute values (since they are the solution to a set of homogeneous equations).
24 The natural frequencies and mode shapes provide a fundamental description of the vibrating
structure.
25 Next we consider the two rigid bar problem illustrated in Fig. 2.11. For this problem, we
must consider equilibrium of the deformed state (rather than undeformed), because of the large
deformation in presence:
P
1
0
0
1
P P 0
1
0
1 P1
0
0
1
0
1
0 P
1
1
0
0
1 1
0 0
1
0
1
0
1
0
1 0
1 0
1
0
1 0
1
1
0
0
1
0000000
1111111
0000000
1111111 C 1
0 0
1
0
1 0
1
0
1
C 0000 C
0000000
1111111
1111
1111111
0000000
0000
1111 0
1 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111 000
111
000
111
1
0000000
1111111
0000
1111 0
1
0
1 0000000
1111111 0000000
1111111 000
111
0000000
1111111
0000
1111 0
1 0000000
1111111
0000000
1111111 0000000
1111111 000
111
0000000
1111111
0000
1111
0000000
1111111
0000
1111 0
1 0000000
1111111 0000000
1111111
0000000
1111111 000
111
000
111
0000000
1111111
0000
1111 0
1 L
0000000
1111111 0000000
1111111 000
111
0000000
1111111
0000
1111 0
1
0
1 0000000
1111111 0000000
1111111 .618
000
111
0000000
1111111
0000
1111 0
1 0000000
1111111 0000000
1111111 000
111
2
0000000
1111111
0000
1111
0000000
1111111
0000
1111 0
1
0
1
0000000
1111111
0000000
1111111 0000000
1111111
1.618
0000000
1111111
000
111
000
111
L 0000000
1111111
0000
1111 0
1 0000000
1111111
2 0P
1 0000000
1111111 000
111
0000000
1111111
0000
1111 0
1 0000000
1111111 0
1 0000000
1111111 000
111
0000000
1111111
0000
1111 0
1 0000000
1111111
0000000
1111111 0
1 0000000
1111111 000
111
0000000
1111111
0000
1111 0
1 0000000
1111111 0
1 0000000
1111111 000
111
0000000
1111111
0000
1111
0000000
1111111
0000
1111 0
1 0000000
1111111 0
1 0000000
1111111
0000000
1111111 000
111
000
111
0000000
1111111
0000
1111 0
1
0
1 0000000
1111111 0
1 0000000
1111111 000
111
0000000
1111111
0000
1111 0
1 0000000
1111111 0
1 0000000
1111111 000
111
B 0000000
1111111
0000
1111 0
1 0000000
1111111 0
1
0
1 0000000
1111111 000
111
0000000
1111111
0000
1111
B 0
1 0000000
1111111
B 0000000
1111111 000
111
k 0000000
1111111
0000
1111 L 0000B
1111
0000
1111 0000000
1111111 000
111
0000
1111 0000
1111 0000
1111 0000
1111
0000
1111 0000
1111 0000
1111
0000
1111
0000
1111 k(2 - )1 1111
0000k( -
1111
0000 )1 0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
1 0000
1111
0000
1111
2 0000
1111
1 0000
1111
1
0000
1111 0000
1111 0000
1111 0000
1111
0000
1111
0000
1111 0000
1111 0000
1111
0000
1111 0000
1111
0000
1111
0000
1111 P 0000
1111 0000
1111 0000
1111
0000
1111 0000
1111
0000
1111 0000
1111 0000
1111
L 0000
1111 0000
1111 0000
1111 0000
1111
0000
1111
0000
1111 0000
1111 0000
1111
0000
1111 0000
1111
0000
1111
0000
1111 0000
1111 0000
1111 0000
1111
0000
1111 0000
1111
0000
1111 0000
1111 0000
1111
0000
1111 0000
1111 0000
1111 0000
1111
0000
1111
0000
1111 0000
1111 0000
1111
0000
1111 0000
1111
0000
1111
0000
1111 0000
1111 0000
1111 0000
1111
0000
1111 0000
1111 0000
1111 0000
1111
Ak 0000
1111
0000
1111 A 1111
0000
k1
0000
1111
0
1
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111 0
1 0000
1111 0000
1111
0
1
0
1
0
1
0
1
0
1
0
1
P
0
1
MB = P L2 k(2 1 ) = 0 (2.63-a)
PL
1 + 2 = 2 (2.63-b)
k
Victor Saouma Finite Elements II; Solid Mechanics
Draft
2.3 Solution of Discrete-System Mathematical models
MA = P L1 + k(2 1 ) k1 = 0
219
(2.63-c)
PL
21 2 = 1 (2.63-d)
k
2 2 + 2 1 = 0 (2.65-c)
3 + 1 = 0
2
(2.65-d)
3 94 3 5
1,2 = = (2.65-e)
2 2
28 We now seek to determine the deformed shape for each of the rst critical loads
3 5
1 = (2.68-a)
# $ 2
2 23 5
1 1 0
= (2.68-b)
1 1 2
3 5 2 0
# $
1+ 5
2 1 1 0
1+ 5
= (2.68-c)
1 1 2 2 0
1.618 1 1 0
= (2.68-d)
1 0.618 2 0
we now arbitrarily set 1 = 1, then 2 = 1/0.618 = 1.618, thus the rst eigenmode is
1 1
= (2.69)
2 1.618
3+ 5
2 = (2.70-a)
# $ 2
2 3+2 5 1 1 0
= (2.70-b)
1 1 3+2 5 2 0
# $
1 5
2 1
1 0
= (2.70-c)
1 1 12 5 2 0
0.618 1 1 0
= (2.70-d)
1 1.618 2 0
we now arbitrarily set 1 = 1, then 2 = 1/1.618 = 0.618, thus the second eigenmode is
1 1
= (2.71)
2 0.618
30 Within the context computer based simulation of a partial dierential equation, we have
three alternatives for a numerical solution
FD
Strong Form
Weak Form
FE
Idealization
Variational Form
Solution
Discretization
Physical Mathematical Discrete Discrete
System Model Model Solution
Solution error
Discretization + solution error
Modeling + discretization+ solution error
Figure 2.12: Idealization, Discretization and Solution of a Numerical Simulation, (Felippa 2000)
31 Two of the three alternative forms for discretization have already been introduced in Finite
Strong Form/Dierential Form, SF/DF which are the Euler Equations corresponding to
the variational formulation. The strong form is represented by a system of ordinary or
partial dierential equations in space and/or time complemented by appropriate boundary
conditions. The strong form is numerically solved by the nite dierence FDM method.
and which can lead to either the strong form (corresponding Euler equations), or to the
weak form. The variational form is solved by the Rayleigh-Ritz method which leads
to a nite element formulation (FEM).
Weak Form, WF: is a weighted integral equation which relaxes the strong form to be
satised on an average sense inside a nite element through the Weighted Residual
method.
32 Note that:
1. The variational method provides a relatively easy way to construct the system of govern-
ing equations. This ease stems from the fact that in the variational formulation scalar
quantities (energies, potentials and so on) are considered rather than vector quantities
(forces, displacements, etc.),
2. Not all problems can be solved by the VF, whereas most can be solved by the WF.
Fluid-structure interaction: FEM for the structure, FDM for the uid.
Structural dynamics: FEM in space, and FDM in time.
33 Given a functional b
(u) = F (x, u, u )dx (2.72)
a
it can be shown that its corresponding Euler equation is given by
F d F
= 0 in a < x < b (2.73)
u dx u
This dierential equation is called the Euler equation associated with and is a necessary
condition for u(x) to extremize .
34 Generalizing for a functional which depends on two eld variables, u = u(x, y) and v =
v(x, y)
= F (x, y, u, v, u,x , u,y , v,x , v,y , , v,yy )dxdy (2.74)
L L
1 1
= M pw dx = (EIw )w pw dx (2.76)
2 2
0 0
F
Solution:
Essential Natural
w = 0 or EIw = M = 0
w = 0 or (EIw ) = V = 0
at x = 0 and x = L
3. The variational form is thus given by
L
! "
(EIw ) + p wdx = 0 (2.79)
0
Euler Eq.
1. Mesh generator
2. Analysis
3. Post-Pocessor
NASTRAN Originally developed by NASA, primarily used by NASA and its contractors.
Original version public domain, later version (McNeal Schwindler) commercial.
SAP Originally developed by Ed. Wilson at Berkeley. First version public domain, later ones
available only for PC (SAP90). NONSAP is the nonlinear version of SAP
ABAQUS Probably the most modern and widely used commercial nite element. (Available
in Bechtel Lab).
FEAP Public domain code listed in Zienkiewicz et al., often used in academia as a base for
extension.
4. Heat transfer
7. Coastal engineering
8. Fluid mechanics
9. Coupled problems
Chapter 3
FUNDAMENTAL RELATIONS
3.1 Introduction
1 Whereas, ideally, a course in Continuum Mechanics should be taken prior to a nite element
course, this is seldom the case. Most often, students have had a graduate course in Advanced
Strength of Materials, which can only provide limited background to a solid nite element
course.
2 Accordingly, this preliminary chapter (mostly extracted from the authors lecture notes in
Continuum Mechanics) will partially remedy for occasional deciencies and will be often ref-
erenced.
3 It should be noted that most, but not all, of the material in this chapter will be subsequently
referenced.
3.1.1 Notation
5 Whereas the Engineering notation may be the simplest and most intuitive one, it often leads
to long and repetitive equations. Alternatively, the tensor will lead to shorter and more compact
forms.
Draft
32 FUNDAMENTAL RELATIONS
6 While working on general relativity, Einstein got tired of writing the summation symbol with
)
its range of summation below and above (such as n=3 i=1 aij bi ) and noted that most of the time
the upper range (n) was equal to the dimension of space (3 for us, 4 for him), and that when
the summation involved a product of two terms, the summation was over a repeated index ) (i
in our example). Hence, he decided that there is no need to include the summation sign if
there was repeated indices (i), and thus any repeated index is a dummy index and is summed
over the range 1 to 3. An index that is not repeated is called free index and assumed to take
a value from 1 to 3.
7 Hence, this so called indicial notation is also referred to Einsteins notation.
8 The following rules dene indicial notation:
1. If there is one letter index, that index goes from i to n (range of the tensor). For instance:
a1
ai = ai = a1 a2 a3 = a i = 1, 3 (3.2)
2
a3
assuming that n = 3.
2. A repeated index will take on all the values of its range, and the resulting tensors summed.
For instance:
a1i xi = a11 x1 + a12 x2 + a13 x3 (3.3)
3. Tensors order:
First order tensor (such as force) has only one free index:
ai = ai = a1 a2 a3 (3.4)
9 Usefulness of the indicial notation is in presenting systems of equations in compact form. For
instance:
x1 = c11 z1 + c12 z2 + c13 z3
xi = cij zj x = c21 z1 + c22 z2 + c23 z3 (3.7)
2
x3 = c31 z1 + c32 z2 + c33 z3
Similarly:
33
A11 = B11 C11 D11 + B11 C12 D12 + B12 C11 D21 + B12 C12 D22
A12 = B11 C11 D11 + B11 C12 D12 + B12 C11 D21 + B12 C12 D22
Aij = Bip Cjq Dpq (3.8)
A = B21 C11 D11 + B21 C12 D12 + B22 C11 D21 + B22 C12 D22
21
A22 = B21 C21 D11 + B21 C22 D12 + B22 C21 D21 + B22 C22 D22
10 In tensor notation the indices are not shown. While cartesian indicial equations apply only to
Cartesian coordinates, expressions in tensor notation are independent of the coordinate system
and apply to other coordinate systems such as cylindrical, curvilinear, etc.
11 In tensor notation, we indicate tensors of order one or greater in boldface letters. It is usually
recommended to use lower case letters for tensors of order one, and upper case letters for tensors
of order 2 and above.
12 We distinguish between tensor notation and matrix notation by using dots and colons between
terms, as in a.b ai bi , (the . indicates a contraction of a pair of repeated indices which appear
in the same order), and A : B Aij Bij . or ij = Cijkl kl is equivalent to = C : .
13In nite element, symmetric second-order tensors are often written as column matrices. This
conversion, and the one of other higher-order tensors into column matrices are called Voigt
notation.
14 The Voigt rule depends on whether the tensor is a kinetic quantity (such as stress) or a
kinematic one (such as strain).
11 1
11 12
22 = = {} (3.9-a)
21 22 2
12 3
11 1
11 12
22 = = {} (3.9-b)
21 22 2
212 3
3.1.2 Tensors
15 We generalize the concept of a vector by introducing the tensor (T), which essentially exists
to operate on vectors v to produce other vectors (or on tensors to produce other tensors!). We
designate this operation by Tv or simply Tv.
16 We hereby adopt the tensor (or dyadic) notation for tensors as linear vector operators
u = Tv or ui = Tij vj (3.10-a)
T
u = vS where S = T (3.10-b)
17 Whereas a tensor is essentially an operator on vectors (or other tensors), it is also a physical
quantity, independent of any particular coordinate system yet specied most conveniently by
referring to an appropriate system of coordinates.
18 Tensors frequently arise as physical entities whose components are the coecients of a linear
23In a contraction, we make two of the indices equal (or in a mixed tensor, we make a subscript
equal to the superscript), thus producing a tensor of order two less than that to which it is
applied. For example:
Tij Tii ; 2 0
ui vj ui vi ; 2 0
..sn A..sm = B.s ;
Amr 4 2
mr r (3.13)
Eij ak Eij ai = cj ; 3 1
Ampr
qs Ampr mp
qr = Bq ; 5 3
24 The outer product of two tensors (not necessarily of the same type or order) is a set of
tensor components obtained simply by writing the components of the two tensors beside each
other with no repeated indices (that is by multiplying each component of one of the tensors by
every component of the other). For example
ai bj = Tij (3.14-a)
i
A Bj.k = C i.k
.j (3.14-b)
vi Tjk = Sijk (3.14-c)
25 The inner product is obtained from an outer product by contraction involving one index
from each tensor. For example
ai bj ai bi (3.15-a)
ai Ejk ai Eik = fk (3.15-b)
Eij Fkm Eij Fjm = Gim (3.15-c)
A i
Bi.k A i
Bi.k =D k
(3.15-d)
body forces: act on the elements of volume or mass inside the body, e.g. gravity, electromag-
netic elds. dF = bdV ol.
surface forces: are contact forces acting on the free body at its bounding surface. Those will
be dened in terms of force per unit area.
28 The surface force per unit area acting on an element dS is called traction or more accurately
stress vector.
tdS = i tx dS + j ty dS + k tz dS (3.17)
S S S S
Most authors limit the term traction to an actual bounding surface of a body, and use the term
stress vector for an imaginary interior surface (even though the state of stress is a tensor and
not a vector).
29 The traction vectors on planes perpendicular to the coordinate axes are particularly useful.
When the vectors acting at a point on three such mutually perpendicular planes is given, the
stress vector at that point on any other arbitrarily inclined plane can be expressed in terms
of the rst set of tractions.
30 A stress, Fig 3.1 is a second order cartesian tensor, ij where the 1st subscript (i) refers to
the direction of outward facing normal, and the second one (j) to the direction of component
force.
11 12 13 t1
= ij = 21 22 23 = t (3.18)
2
31 32 33 t3
31 In fact the nine rectangular components ij of turn out to be the three sets of three vector
components (11 , 12 , 13 ), (21 , 22 , 23 ), (31 , 32 , 33 ) which correspond to the three tractions
t1 , t2 and t3 which are acting on the x1 , x2 and x3 faces (It should be noted that those tractions
are not necesarily normal to the faces, and they can be decomposed into a normal and shear
traction if need be). In other words, stresses are nothing else than the components of tractions
(stress vector), Fig. 3.1.
32 The state of stress at a point cannot be specied entirely by a single vector with three
X3
FUNDAMENTAL RELATIONS
X3
V3
33
32
t3 V
23
31
V2
X2
t2
13 22 V1
X1
21
X2
(Components of a vector are scalars)
12
t1
11
33 Let us now consider the problem of determining the traction acting on the surface of an
oblique plane (characterized by its normal n) in terms of the known tractions normal to the
three principal axis, t1 , t2 and t3 . This will be done through the so-called Cauchys tetrahedron
shown in Fig. 3.2, and will be obtained without any assumption of equilibrium and it will apply
X2
B S3
-t *
1
*
-t 3
S
1 t*n S
n
h N A
O X1
C
-t 2 S2
*
X3
V
*
*
b
ponents of tn are
37
tn1 = 11 n1 + 21 n2 + 31 n3
tn2 = 12 n1 + 22 n2 + 32 n3
tn3 = 13 n1 + 23 n2 + 33 n3 (3.19)
Indicial notation tni = ji nj
Tensor notation tn = n = T n
35 We have thus established that the nine components ij are components of the second order
tensor, Cauchys stress tensor.
We seek to determine the traction (or stress vector) t passing through P and parallel to the
plane ABC where A(4, 0, 0), B(0, 2, 0) and C(0, 0, 6).
Solution:
The vector normal to the plane can be found by taking the cross products of vectors AB and
AC:
e 1 e2 e3
N = ABAC = 4 2 0 (3.21-a)
4 0 6
= 12e1 + 24e2 + 8e3 (3.21-b)
and thus t = 97 e1 + 57 e2 + 10
7 e3
3.2.2 Kinematic
Spatial Coordinates (x1 , x2 , x3 ) dened in the deformed coordinate system. This gigives rise
to the Eulerian coordinate system
1. If both the displacement gradients and the displacements themselves are small, then
Xj xj and thus the Eulerian and the Lagrangian innitesimal strain tensors may be
ui ui
.
taken as equal Eij = Eij
2. If the displacement gradients are small, but the displacements are large, we should use
the Eulerian innitesimal representation.
3. If the displacements gradients are large, but the displacements are small, use the La-
grangian nite strain representation.
4. If both the displacement gradients and the displacements are large, use the Eulerian nite
strain representation.
or:
# $
u1 1 u1 2 u2 2 u3 2
E11 = + + + (3.25-a)
X1 2 X1 X1 X1
1 u1 u2 1 u1 u1 u2 u2 u3 u3
E12 = + + + + (3.25-b)
2 X2 X1 2 X1 X2 X1 X2 X1 X2
= (3.25-c)
38 If large deformation is accounted for (such as in buckling), the Eulerian nite strains are:
# 2 2 2 $
u 1 u v w
xx = + + + (3.26)
x 2 x x x
# 2 $
v 1 u 2 v w 2
yy = + + + (3.27)
y 2 y y y
# $
w 1 u 2 v 2 w 2
zz = + + + (3.28)
z 2 z z z
1 v u u u v v w w
xy = + + + + (3.29)
2 x y x y x y x y
1 w u u u v v w w
xz = + + + + (3.30)
2 x z x z x z x z
1 w v u u v v w w
yz = + + + + (3.31)
2 y z y z y z y z
or
39
1
ij = (ui,j + uj,i + uk,i uk,j ) (3.32)
2
ij = 2ij (i = j) (3.33)
2. If the strains are given, then these strain-displacements provide a system of (6) nonlinear
partial dierential equation in terms of the unknown displacements (3).
4. The strains have been expressed in terms of the coordinates x, y, z in the undeformed state,
i.e. in the Lagrangian coordinate which is the preferred one in structural mechanics.
6. In most cases the deformations are small enough for the quadratic term to be dropped,
the resulting equations reduce to
u
xx = (3.34)
x
v
yy = (3.35)
y
w
zz = (3.36)
z
v u
xy = + (3.37)
x y
w u
xz = + (3.38)
x z
w v
yz = + (3.39)
y z
or
1
ij = (ui,k + uk,i ) (3.40)
2
39 In nite element, the strain is often expressed through the linear operator L
= Lu (3.41)
or
FUNDAMENTAL RELATIONS
xx 0 0
0
x
yy
0
0
y
ux
zz 0 z
= uy
xy
y 0 (3.42)
x
uz
z 0
xz
x
u
yz 0
z
y
L
40 In this section, we will derive dierential equations governing the way stress and deformation
vary at a point and with time. They will apply to any continuous medium, and yet we will
not have enough equations to determine unknown tensor eld. For that we need to wait for
the next section where constitututive laws relating stress and strain will be introduced. Only
with constitutive equations and boundary and initial conditions would we be able to obtain a
well dened mathematical problem to solve for the stress and deformation distribution or the
displacement or velocity elds.
41 In this section we shall summarize the dierential equations which express locally the con-
tion law establishes a balance of a scalar or tensorial quantity in voulme V bounded by a surface
S. In its most general form, such a law may be expressed as
d
AdV + dS = AdV (3.43)
dt V
S V
Rate of variation Exchange by Diusion Source
where A is the volumetric density of the quantity of interest (mass, linear momentum, energy,
...) a, A is the rate of volumetric density of what is provided from the outside, and is the
rate of surface density of what is lost through the surface S of V and will be a function of the
normal to the surface n.
44 Hence, we read the previous equation as: The input quantity (provided by the right hand
side) is equal to what is lost across the boundary, and to modify A which is the quantity of
interest. The dimensions of various quantities are given by
45Hence this section will apply the previous conservation law to mass, momentum, and energy.
The resulting dierential equations will provide additional interesting relation with regard to
n vn dt
vdt
dS
d vi d
+ = 0 or + v = 0 (3.51)
dt xi dt
The vector form is independent of any choice of coordinates. This equation shows that the
divergence of the velocity vector eld equals (1/)(d/dt) and measures the rate of ow of
material away from the particle and is equal to the unit rate of decrease of density in the
neighborhood of the particle.
50 If the material is incompressible, so that the density in the neighborhood of each material
particle remains constant as it moves, then the continuity equation takes the simpler form
vi
= 0 or v = 0 (3.52)
xi
51 The momentum principle states that the time rate of change of the total momentum of a
given set of particles equals the vector sum of all external forces acting on the particles of the
set, provided Newtons Third Law applies. The continuum form of this principle is a basic
postulate of continuum mechanics.
d
tdS + bdV = vdV (3.53)
S V dt V
Then we substitute ti = Tij nj and apply the divergence theorm (Appendix A, (Schey 1973)) to
obtain
Tij dvi
+ bi dV = dV (3.54-a)
V x j V dt
Tij dvi
+ bi dV = 0 (3.54-b)
V xj dt
or for an arbitrary volume
Tij dvi dv
+ bi = or T + b = (3.55)
xj dt dt
53 We note that these equations could also have been derived from the free body diagram
shown in Fig. 3.4 with the assumption of equilibrium (via Newtons second law) considering
an innitesimal element of dimensions dx1 dx2 dx3 . Writing the summation of forces, will
yield
Tij,j + bi = 0 (3.57)
xx d x
xx+
xx x
dy
xy x
xy xy+ d
x
yx
yy
dx
54 The rst principle of thermodynamics relates the work done on a (closed) system and the
heat transfer into the system to the change in energy of the system. We shall assume that
the only energy transfers to the system are by mechanical work done on the system by surface
traction and body forces, by heat transfer through the boundary.
55 If mechanical quantities only are considered, the principle of conservation of energy for
the continuum may be derived directly from the equation of motion given by Eq. 3.55. This is
accomplished by taking the integral over the volume V of the scalar product between Eq. 3.55
and the velocity vi .
dvi
vi Tji,j dV + bi vi dV = vi dV (3.58)
V V V dt
Applying the divergence theorem,
dK dU dW
+ = +Q (3.59)
dt dt dt
this equation relates the time rate of change of total mechanical energy of the continuum on
the left side to the rate of work done by the surface and body forces on the right hand side.
56 If both mechanical and non mechanical energies are to be considered, the rst principle states
that the time rate of change of the kinetic plus the internal energy is equal to the sum of the
rate of work plus all other energies supplied to, or removed from the continuum per unit time
(heat, chemical, electromagnetic, etc.).
57 For a thermomechanical continuum, it is customary to express the time rate of change of
where u is the internal energy per unit mass or specic internal energy. We note that U
appears only as a dierential in the rst principle, hence if we really need to evaluate this
quantity, we need to have a reference value for which U will be null. The dimension of U is one
of energy dim U = M L2 T 2 , and the SI unit is the Joule, similarly dim u = L2 T 2 with the
SI unit of Joule/Kg.
3.2.4.1 General 3D
ij = Dijkl kl i, j, k, l = 1, 2, 3 (3.61)
59The (fourth order) tensor of elastic constants Dijkl'has 81((34 ) components however, due to
the symmetry of both and , there are at most 36 9(91)2 distinct elastic terms.
60 For the purpose of writing Hookes Law, the double indexed system is often replaced by a
simple indexed system with a range of six:
62 =36
k = Dkm m k, m = 1, 2, 3, 4, 5, 6 (3.62)
1 +
= ; = (3.65)
E (3 + 2) 2( + )
E E
= ; = G = (3.66)
(1 + )(1 2) 2(1 + )
and
FUNDAMENTAL RELATIONS
1 1 1
a11 = ; a12 = ; a13 = ; a33 = ; a44 = (3.73)
E E E E
where E is the Youngs modulus in the plane of isotropy and E the one in the plane normal
to it. corresponds to the transverse contraction in the plane of isotropy when tension is
applied in the plane; corresponding to the transverse contraction in the plane of isotropy
when tension is applied normal to the plane; corresponding to the shear moduli for the plane
of isotropy and any plane normal to it, and is shear moduli for the plane of isotropy.
68 Often times one can make simplifying assumptions to reduce a 3D problem into a 2D one.
3.2.4.3.1 Plane Strain 69 For problems involving a long body in the z direction with no
variation in load or geometry, then zz = yz = xz = xz = yz = 0. Thus, replacing into Eq.
3.71 we obtain
(1 ) 0
xx
xx
yy E (1 ) 0
=
(1 + )(1 2) 0 yy (3.74)
zz
xy
12
xy 0 0 2
3.2.4.3.3 Plane Stress 72 If the longitudinal dimension in z direction is much smaller than
in the x and y directions, then yz = xz = zz = xz = yz = 0 throughout the thickness.
xx 1
1 0 xx
yy = 1 0 (3.77-a)
1 2 1 yy
xy 0 0 2 xy
1
zz = (xx + yy ) (3.77-b)
1
73 In porous material, the water pressure is transmitted to the structure as a body force of
magnitude
p p
bx = by = (3.78)
x y
ij = ij + mij p mT = [ 1, 1, 0 ] (3.79)
i.e simply removing the hydrostatic pressure component from the total stress.
81 It must be noted that the pore pressures p being considered in this discussion and throughout
the remainder of this course are the steady state pore pressures; excess pore pressures resulting
from dilatant behavior in the skeleton of the material are not considered.
82 Step by step generalization of the eld equations:
0ij = T ij (3.81)
ij = T Dijkl kl (3.82)
Pore pressure
0
ij = p ij (3.86)
p the pore pressure dened using the compression positive sign convention (the minus
sign corrects the discrepancy in the sign conventions)1
t i = ti + p ni + T Dijkl ni kl (3.88)
83 The eective stress principle for the case of combined thermo- and poro-elasticity is
ij = ij ij 0
+ ij (3.89)
ij
84 The general stress-strain relationship obtained by substituting the constutive law into the
319
85 When thermal strains and pore pressures are considered in combination the constitutive law
86 The equilibrium equation and natural boundary conditions, respectively, can be rewritten in
terms of the eective stresses ij
:
ij,j + b i = 0 Equilibrium
(3.92)
t : ij nj t i = 0 Natural B.C.
where
b i = bi p,i Dijkl T,i kl
(3.93)
t i = ti + p ni + T Dijkl ni kl
87 Scalar eld problems are encountered in almost all branches of engineering and physics. Most
of them can be viewed as special forms of the general Helmholtz equation given by
kx + ky + kz + Q = c (3.94)
x x y y z z t
div(Dr) + Q = 0
Equation D Q q Constitutive law
Heat Flow Temperature Thermal Heat Heat Fourrier
T conductivity supply/sink ux q = DrT
Fluid ow through Piezometric Permeability Fluid Volume Darcy
porous media head h Coecients supply ux q = Dr
Diusion ion Permeability Ion Ion Fick
concentration coecients supply ux q = Dr
1
Saint-Venant Prandtls
Rate of Hooke
Torsion stress function twist 2
Conduction: takes place when a temperature gradient exists within a material and is governed
by Fouriers Law, Fig. 3.5 on q :
T
qx = kx (3.95)
x
T
qy = ky (3.96)
y
where T = T (x, y) is the temperature eld in the medium, qx and qy are the componenets
of the heat ux (W/m2 or Btu/h-ft2 ), k is the thermal conductivity (W/m.o C or Btu/h-
ft-o F) and T T
x , y are the temperature gradients along the x and y respectively. Note
that heat ows from hot to cool zones, hence the negative sign.
Convection: heat transfer takes place when a material is exposed to a moving uid which is
at dierent temperature. It is governed by the Newtons Law of Cooling
q = h(T T ) on c (3.97)
where q is the convective heat ux, h is the convection heat transfer coecient or lm
coecient (W/m2 .o C or Btu/h-ft2 .o F). It depends on various factors, such as whether
convection is natural or forced, laminar or turbulent ow, type of uid, and geometry of
the body; T and T are the surface and uid temperature, respectively. This mode is
considered as part of the boundary condition.
Radiation: is the energy transferred between two separated bodies at dierent temperatures
by means of electromagnetic waves. The fundamental law is the Stefan-Boltmans Law of
Thermal Radiation for black bodies in which the ux is proportional to the fourth power
of the absolute temperature, which causes the problem to be nonlinear. This mode will
not be covered.
90Consider a solid through which there is a ow q of some quantity, mass, heat, chemical, etc...
The rate of transfer per unit area is q which is nothing else than the ux.
91 The direction of ow (i.e. ux) is in the direction of maximum potential (such as temper-
ature, piezometric head, or ion concentration) decrease (Fourrier, Darcy, Fick...). Hence the
ux is equal to the gradient of the scalar quantity.
x
q
x
q= qy = D = D (3.98)
y
qz
z
Anisotropic
kxx kxy kxz
D = kyx kyy kyz (3.100)
kzx kzy kzz
Orthotropic
kxx 0 0
D = 0 kyy 0 (3.101)
0 0 kzz
Note that for ow through porous media, Darcys equation is only valid for laminar ow.
92 If we consider a unit thickness, 2D dierential body of dimensions dx by dy, Fig. 3.6 then
1. Rate of heat generation/sink is
I2 = Qdxdy (3.102)
2. Heat ux across the boundary of the element is shown in Fig. ?? (note similarity with
equilibrium equation)
qx qy qx qy
I1 = qx + dx qx dx dy + qy + dy qy dy dx = dxdy + dydx
x y x y
(3.103)
q + qy dy
y y
qx
qx qx +
x dx
Q dy
qy
dx
93 From the rst law of thermodynamics, energy produced I2 plus the net energy across the
94 The amount of ow per unit time into an element of volume and surface is
I1 = q(n)d = D.nd (3.106)
where n is the unit exterior normal to , Fig. 3.7
95 Using the divergence theorem
vnd = div vd (3.107)
97
323
Finally, we dene the specic heat of a solid c as the amount of heat required to raise a unit
mass by one degree. Thus if is a temperature change which occurs in a mass m over a time
t, then the corresponding amount of heat that was added must have been cm, or
I3 = cd (3.110)
energy across the boundary I1 must be equal to the energy absorbed I3 , thus
I1 + I2 I3 = 0 (3.111-a)
div (D) + Q c d = 0 (3.111-b)
t
qx qy
+ + Q = c (3.114)
x y t
1. Note the similarity between this last equation, and the equation of equilibrium
xx xy 2 ux
+ + bx = m 2 (3.115-a)
x y t
yy xy 2 uy
+ + by = m 2 (3.115-b)
y x t
2. For steady state problems, the previous equation does not depend on t, and for 2D
problems, it reduces to
kx + ky +Q=0 (3.116)
x x y y
2 2 2 Q
2
+ 2 + 2 = (3.117)
x y z k
which is Poissons equation in 3D.
2 2 2
+ 2 + 2 =0 (3.118)
x2 y z
which is an Elliptic (or Laplace) equation. Solutions of Laplace equations are termed
harmonic functions (right hand side is zero) which is why Eq. 3.116 is refered to as the
quasi-harmonic equation.
Natural: Flux
99 The analogy between scalar and vector problems is shown in Table 3.2.
100To graphically illustrate the inter-relationship between eld equations and variables, Tonti
diagrams, Fig. 3.8 have been used.
Essential
Specified B.C.
Primary Variable Primary Variable Field Source Function
(EBC)
Equations
Kinematic Balance
Equation Equations
(KE) Constitutive (BE)
Equations Natural
Intermediate (CE) Flux B.C. Specified
Variable Variable (NBC) Flux Variable
Essential B.C.
ui : u
Body Forces State variable
bi ui
Balance Kinematics
LT + b = 0 " = Lu
Flux variable Constitutive Rel. Interm. variable
ij ij = Dijkl kl ij
Natural B.C.
t i : t
Balance State variable
r.q = Q T
Flux variable Constitutive Rel. Interm. variable
q q = Dg g=rT
Natural B.C.
qn : q c
Scalar Vector
FUNDAMENTAL RELATIONS
101 However, we here expand on the standard Tonti diagram to graphically show the fundamen-
Victor Saouma
Physical Conservation Prin- State Vriable Flux Material Con- Source Constitutive
Problem ciple stants Equation
Deformation of Equilibrium of forces Displacement Stress or strain Youngs Mod- Body force or Hookes law
an elastic body or forces ulus, Poissons traction
ratio
Electric net- Equilibrium of cur- Voltage or in- Electric ux Electric conduc- External elec- Kirchhos law
work rents tensity tivity tric charge
3.4 Summary and Tonti Diagrams
Torsion Conservation of poten- Stress or warp- Rate of twist Shear -2 angle of Hookes law
tial energy ing function twist
Heat transfer Conservation of energy Temperature Heat ux Thermal conduc- Internal or ex- Fourriers law
tivity ternal heat
Fluid ow Conservation of mo- Velocity Shear stress Viscosity Body forces Stokes law
mentum
Flow through Conservation of mass Hydraulic head Flow rate Permeability Fluid sources Darcys law
porous media
Electrostatics Conservation of elec- Electric poten- Electric ux Permitivity Charge Coulombs law
tric ux tial
Magnetostatic Conservation of mag- Magnetic Magnetic ux Magnetic perme- Current Maxwells law
netic potential potential ability
Chapter 4
MESH GENERATION
4.1 Introduction
1 Finite element mesh generation is now an integral part of a nite element analysis. With the
increased computational capabilities, increasingly more complex structures are being analysed.
Those structures must be discretized.
2 The task is one of developing a mathematical model (discretization or tessalation) of a con-
tinuum model. This is not only necessary in nite elment analysis, but in computer graph-
ics/rendering also.
3 In computer graphics, we focus on the boundary representation, and assign colors and shades
on the basis of light source and outward normal direction of the polygon.
4 Hence, in the most general case, meshing can be dened as the process of breaking up a physi-
cal domain into smaller sub-domains (elements) in order to facilitate the numerical solution of a
partial dierential equation. Surface domains may be subdivided into triangle or quadrilateral
shapes, while volumes may be subdivided primarily into tetrahedra or hexahedra shapes. The
shape and distribution of the elements is ideally dened by automatic meshing algorithms.
2. Sub-domain removal. Elements are gradually removed from the domain, one ata time,
until the whole domain is decomposed int nite elements.
3. Recursive subdivision. The domain is broken into simpler parts until the individual parts
are single elements or simple regions, that can be meshed directly, for instance by the
conformal mapping algorithm.
4.2 Triangulation
MESH GENERATION
5 The concept of Voronoi diagrams rst appeared in works of Descartes as early as 1644.
Descartes used Voronoi-like diagrams to show the disposition of matter in the solar system and
its environs.
6 The rst man who studied the Voronoi diagram as a concept was a German mathematician
G. L. Dirichlet. He studied the two- and three dimensional case and that is why this concept
is also known as Dirichlet tessellation. However it is much better known as a Voronoi diagram
because another German mathematician M. G. Voronoi in 1908 studied the concept and dened
it for a more general n-dimensional case.
7 Very soon after it was dened by Voronoi it was developed independently in other areas like
Va Vb
a
b
c
Vc
1 2 3 4 4f
6 6f
Voronoi
Delaunay
9 Given a nite set of poits in the plane, the idea is to assign to each point a region of inuence
in such a way that the regions decompose the plane, (Voronoi 1907).
Voronoi region of p S as the set of points x R that are at least as close to p as to any
other points in S: * +
Vp = x R2 | ||x p|| x 1 , q S (4.1)
Each point x R 2 has at least one nearest point in S, so it lies in at least one Voronoi region.
Two Voronoi regions lie on opposite sides of the perpendicular bisector separating the two
generating points.
11 The dual of the Voronoi diagram is obtained by drawing straight Delaunay edges connecting
points p, q S if and only if their Voronoi regions intersect along a common line segment. Thus
in general, the Delaunay edges decompose the convex hull of S into triangular regions which
are referred to as Delaunay triangles, (Delaunay 1934).
12 Using Eulers relation, it can be shown that a planar graph with n 3 vertices has at most
3n 6 edges and at most 2n 4 faces. THe same bounds hold for the number of Delaunay
edges and triangles.
13Each Voronoi vertex u = Va Vb Vc is the center of a circle with radius = u a =
u b = u c . The circle is empty because it encloses no point of S.
14 Additional detailed information on Voronoi tesselation, (Okabe, Boots, Sugihara and Chiu
15 In order to discretize the continuum into a nite element mesh, rst key geometrical infor-
Vertices: with nodal coordinates, and approximate desired element size in the immediate
vicinity (thus describing the mesh density).
Edges: which connect vertices. Those can be either linear segments, polylines, or curves.
16 Associated with surfaces (2D0, or volumes (3D) are dierent material properties.
17 Examples of the hierarchical boundary denition is shown in Fig. 4.2 and 4.3.
18 Once the boundary has been dened, we need to insert internal nodes at a spacing which
respect the required mesh density. There are numerous techniques to insert those internal
nodes. We present one approach, FIg. 4.4
1. Decompose the region into a disjoint ensemble of subregions with equal mesh density.
2. Shrink the mesh to avoid elements near the boundary with very acute interior angles.
3. Starting with the rst zone, circumscribe it by the smallest possible rectangle.
Zone II, r2
r
Generation of internal nodes
in zone I within shrunk boundary
Disk
r
5. Use a random number generator to randomly generate one interior node in each square.
A disk of radius r centered at each node is used to test that no other surrounding nodes
are enclosed in the disk. If so, the node in question is regenerated.
2. Check lengths of the edges. If an edge does not satisfy the prescribed size r, a new node
is inserted in the center of the edge. The prescribed size is interpolated between those of
the vertices at each end of the edge.
3. Smoothen the mesh to ensure that all generated elements have a satisfactory aspect ratio.
21 It should be noted that recent algorithms, which can generate quadrilateral elements out of
Chapter 5
VARIATIONAL and
RAYLEIGH-RITZ METHODS
1 A Multield variational principle is one that has more than one master eld (or state
variable), that is more than one unknown eld is subject to independent variations.
2 In linear elastostatics, we can have displacement, u, strains , or stress as potential can-
didates for master elds. Hence seven combinations are possible, (Felippa 2000), Table 5.1.
3 In this course, we shall focus on only the Total Potential Energy, and the Hu-Washizu varia-
tional principles.
u Name
Single Field
Y Total Potential Energy
Y Total Complementary Potential Energy
Y No name
Two Fields
Y Y Hellinger-Reissner
Y Y de Veubeke
Y Y No name
Three Fields
Y Y Y Hu-Washizu
the functional for the general form of the potential energy variational principle is obtained
1
= D d
T
D 0 d +
T
0 d
T
u bd
T
uT td uP
2 (5.2)
t
U We
5 A variational statement is obtained by taking the rst variation of and setting this scalar
(Eq. 3.41)
= Lu (5.4)
where L is a linear dierential operator and u is the displacement vector dened in Eq. 3.42:
x 0 0
0
0
y
0
0 z
L= (5.5)
y x 0
z 0
x
0 z y
8 Since the dierential operator L is linear, the variation of the strains can be expressed in
= (Lu) = Lu (5.6)
1. In terms of strains: which is more suitable for the derivation of the corresponding Euler
53
equations,
= (Lu) D d
T T
(Lu) D 0 d + (Lu)T 0 d
U
uT bd uT td = 0 (5.7)
t
We
2. In terms of displacements: (using Equation 5.7) which is more suitable for the subse-
quent discretization.
= (Lu) D(Lu)d
T T
(Lu) D 0 d + (Lu)T 0 d
U
uT bd uT td = 0 (5.8)
t
We
10 To obtain the Euler equations for the general form of the potential energy variational principle
the volume integrals dening the virtual strain energy U in Equation 5.7 must be integrated by
parts in order to convert the variation of the strains (Lu) into a variation of the displacements
u.
11 Integration by parts of these integrals using Greens theorem (Eq. 1.46)
,
S R
R d = Sd + RSnx d (5.9)
x x
yields
,
T
(Lu) D d = u G(D )d
T
uT LT (D )d (5.10-a)
,
(Lu)T D 0 d = uT G(D 0 )d uT LT (D 0 )d (5.10-b)
,
(Lu) 0 d =
T
u G 0 d uT LT 0 d
T
(5.10-c)
where G is a transformation matrix containing the direction cosines for a unit normal vector
such that the surface tractions t are dened as t = G and the surface integrals are over the
entire surface of the body .
12 Substituting Equation 5.10-a into Equation 5.7, the variational statement becomes
= uT {LT [D( 0 ) + 0 ] +b}d
u u
u b
^u ^
t
Figure 5.1: Tonti Diagram for the Total Potential Energy, (Cervenka, J. 1994)
+ uT {G [D( 0 ) + 0 ] t}d = 0 (5.11)
13 Since u is arbitrary the expressions in the integrands within the braces must both be equal
to zero for to be equal to zero. Recognizing that the stress-strain relationship appears in
both the volume and surface integrals, the Euler equations are
the total potential energy functional by performing the operations just presented in reverse
order.
16 The Tonti diagram for the TPE is shown in Fig. 5.2. In this diagram, strong connections are
shown by solid lines, weak connections by spring-like symbols, boxes with solid lines denote the
primary unknown eld, variables inside dashed boxes are internally derived elds, and shaded
boxes indicate prescribed elds.
17In dynamics, we dene the Lagrangian function as the dierence between the kinetic and
the potential energies
LK (5.14)
and Hamiltons principle states that The motion of a particle acted on by conservative forces
between two arbitrary instants of time t1 and t2 is such that the line integral over the Lagrangian
function is an extremum for the path motion.
t2
I Ldt (5.15)
t1
is an extremum.
55
18If the path can be dened in terms of generalized coordinates qi (i = 1, 2, 3), then it can be
shown that
t2
I = L(q1 , q2 , q3 , q1 , q2 , q3 )dt = 0 (5.16-a)
t1
t2 3
L d L
= qi dt (5.16-b)
t1 i=1 qi dt qi
19 If all qi are linearly independent (i.e. no constraints among them), the variations qi are
independent of t, except qi = 0 at t1 and t2 . Therefore, the coecients of q1 , q2 , and q3
vanish separately yielding the Lagrangian equations of motion, also known as the Euler-
Lagrange equations.
L d L
= 0, i = 1, 2, 3 (5.17)
qi dt qi
where K is the kinetic energy, the potential energy, and Wnc the work done by nonconservative
forces acting on the system (including damping).
It should be noted that TPE is a special case of Hamiltons principle in the absence of
kinetic energy.
The Kinetic energy is given by
u u
K= . d (5.20)
2 t t
A conservative force is one for which the sum of the potential and kinetic energies is con-
served).
For a uniform cross section bar of length L, cross sectional area A, Youngs modulus E, and
mass density xed at one end, and connected to a rigid support at the other by a spring with
stiness k
k
A, E
L
1. Show that the kinetic energy and the strain energy are given by
u 2
L
A
K = dx (5.21)
t0 2
L
EA u 2 k
U = d + [u(L)]2 (5.22)
0 2 x 2
2. For Wnc = 0 derive an expression for the rst variation of the Hamiltons functional H
3. If we are interested in determining the periodic motion, which has the form
where is the frequency of natural vibration, and u0 (x) is the amplitude. Show that
# 2 $
L
1 du0 k
A 2 (u0 )2 EA dx [u0 (L)]2 = 0 (5.24)
0 2 dx 2
x u0 kL 2 L2
5. Rewrite Eq. 5.24 in terms of x = L, u= L, = EA , = E
6. Show that if this problem was to be solved by the Rayleigh-Ritz method with
u = c1 x + c2 x2 (5.27)
7. Solve for 1 and 2 , and compare with the exact solution given by
+ tan = 0 (5.29)
Solution:
1. Kinetic energy
L
u 2
A u 2
K = d = dx (5.30)
2
t 0 2 t
L
E u 2 EA u 2 k
U = d = dx + [u(L)]2 (5.31)
2 x 0 2 x 2
2. For Wnc = 0
57
t2
H = (K )dt (5.32)
t1
t2 L # 2 2 $
1 u u
= A EA dx k[u(L)]2 dt (5.33)
2 t1 0 t x
3. If we adopt
u(x, t) = u0 (x)eit (5.34)
where is the frequency of natural vibration, and u0 (x) is the amplitude. Then substi-
tuting, we obtain
# $
L
1 u0 2 k
A [u0 (x)] EA
2 2
dx [u0 (L)]2 = 0 (5.35)
0 2 x 2
t2
where exp2it dt being nonzero is factored out.
t1
but
L
L
u0 u0 u0 L
u0
EA dx = EA
u0 EA u0 dx (5.38)
x x x x x
0
v 0
v
u v u 0 u
Substituting
L
u0 u0
A [u0 (x)] EA
2
dx ku0 (L)u0 (L) (5.39)
0 x x
L L
d u0 du0
= A[u0 (x)] + EA EA u0 ku0 (L)u0 (L) = 0(5.40)
0 dx x dx
0
0
or
d du0
EA + A 2 u0 = 0 for 0 < x < L (5.41)
dx dx
du0
EA + ku0 = 0 for x = L (5.42)
dx
5. Substituting x = x
VARIATIONAL and RAYLEIGH-RITZ METHODS
u = c1 x + c2 x2 (5.49)
Collecting the coecients of c1 and c2 and setting them to zero separately, we have
1
! "
c1 : (c1 x + c2 x2 )x (c1 + 2c2 x) dx (c1 + c2 ) = 0 (5.52)
0
1
! "
c2 : (c1 x + c2 x2 )x2 (c1 + 2c2 x)2x dx (c1 + c2 ) = 0 (5.53)
0
or
152 640 + 2400 = 0 (5.56)
which correspond to
2.038 E 6.206 E
1 = 2 = (5.58)
L L
+ tan = 0 (5.59)
the displacements, strains, and stresses are treated as independent elds (as opposed
to only the displacement in the total potential energy principle).
21 Naturally, the two additional eld variables, with respect to the TPE variational principle,
appear not only in the functional, but also in the discretized system of equations. Consequently,
for a domain with a given discretization the discrete system of equations derived from the HW
variational principle will be much larger than the discrete system of equations derived from the
TPE variational principle.
22 With the increased number of equations, signicant improvements in accuracy can be ob-
served for the solution obtained from the discrete form of the HW variational principle compared
to the solution obtained from the discrete form of the TPE variational principle for the same
discretization. This means that coarse discretizations can be used with the discrete form of the
HW variational principle to obtain the same degree of accuracy that would be observed with
much ner discretizations using the TPE variational principle.
23 The functional for the HW variational principle is derived from the functional for the TPE
Lu = 0 (5.61)
25
VARIATIONAL and RAYLEIGH-RITZ METHODS
Where is the Lagrange multiplier and to be consistent with the integrals in the TPE functional
(i.e. Equation 5.2) the Lagrange multiplier must have the units of stress. Since this is the case,
will be used for the Lagrange multiplier instead of the more typical such that the physical
meaning of the Lagrange multiplier is more apparent.
27 The functional for the HW variational principle thus becomes
1
HW = D d
T
D 0 d +
T
T 0 d
2
U (5.63)
uT b d uT t d + T (L u ) d
t
We Constraint
28 A variational statement is obtained by taking the rst variation of the functional and setting
Note that the 4th and 7th term were added and cancell each others, and that we are not using
Eq. 5.6 in this formulation.
30 Since u, , and are independent eld variables, terms involving u, , and must add up
to zero individually and are, therefore grouped together to form three separate variational state-
ments (analogous to the method of separation of variables in the solution of partial dierential
equations)
(Lu)T d uT b d uT t d = 0 (5.65)
t
T [D( 0 ) + 0 ] d = 0 (5.66)
T (L u ) d = 0 (5.67)
31 To obtain the corresponding Euler equations for the general form of the HW variational
511
principle the volume integral in Equation 5.65 containing the variation of the strains (Lu)
dened in terms of the displacements u must be integrated by parts using Greens theorem in
order to obtain a form of the variational statement in terms of the variation of the displacements
u.
32 Integration by parts (Eq. 1.39) of this integral yields
,
(L u) d =
T
u G d
T
uT LT d (5.68)
where G is a transformation matrix containing the direction cosines for a unit normal vector
such that the surface tractions t are dened as t = G and the surface integral is over the
entire surface of the body .
33 Substituting Equation 5.68 into Equation 5.65, the rst variational statement becomes
u (L + b) d +
T T
uT (G t) d = 0 (5.69)
t
Since u is arbitrary the expressions in the integrands within the parentheses must both be
equal to zero for the sum of the integrals to be equal to zero.
34 Likewise, and are also arbitrary and the expressions within the braces in the second
variational statement (i.e. Equation 5.66) and within the parentheses in the third variational
statement (i.e. Equation 5.67) must both be equal to zero for the integral to be equal to zero.
35 The Euler equations for the HW functional are
where the rst Euler equation is the equilibrium equation; the second Euler equation is the
stress-strain relationship; the third Euler equation is the strain-displacement equation; and the
fourth Euler equation denes the natural boundary conditions.
36 The natural boundary conditions are dened on t rather than because both the applied
surface tractions t and the matrix-vector product G are identically zero outside t . Starting
from the Euler equations, it is possible to derive the HW functional by performing the operations
just presented in reverse order.
37 This last set of four Euler equations, should be compared with the two (Eq. 5.12 and 5.13)
obtained from the original TPE. The additional two equations bring into play stress-strain and
strain displacement. Also, whereas the original formulation (Eq. 5.12 and 5.13) was in terms
of the displacement only (u), the Hu-Washizu formulation is in terms of three independent
variables (u, and ), Table 5.2.
38 The Tonti diagram for the HW is shown in Fig. ??.
TPE HW
Euler Equations
Equilibrium LT + b = 0 Y Y
Stress-Strain D ( 0 ) + 0 = 0 N Y
Strain-Displacement Lu = 0 N Y
Natural B.C. G t = 0 t Y Y
Variables
Displacement u Y Y
Strain N Y
Stress N Y
Table 5.2: Comparison Between Total Potential Energy and Hu-Washizu Formulations
u u
u
^
u
^
t
39 In the principle of virtual displacement (or minimum total potential energy), the Euler equa-
tions are the equilibrium equations. The Euler equations are usually in the form of dierential
equations that are not always solvable by exact methods of solutions.
40Indeed we do have available a number of approximate methods such as the nite dierence
but an alternative approach is to bypass entirely the Euler equations and start directly from a
variational statement of the problem to the solution of the Euler equation.
41 Such a direct method was rst proposed by Lord Rayleigh, and then independently general-
ized by Ritz.
42 Hence, this section will briey review the Rayleigh-Ritz method (which solves the variational
problem, but which does not require a topological discretization (i.e. nite element). The
similarity between those two approaches should by now be apparent. In both cases we are
dealing with an approximate method.
43 Application of the principle of total potential energy (or virtual displacement) requires an
1. Unknown polynomial coecients most appropriate for continuous systems, and the
Rayleigh-Ritz method
y = a1 + a2 x + a3 x2 + a4 x3 (5.74)
A major drawback of this approach, is that the coecients have no physical meaning.
2. Unknown nodal deformations most appropriate for discrete systems and Potential En-
ergy based formulations
y = u = N1 u1 + N2 u2 + . . . + Nn un (5.75)
44 This chapter will focus on the rst type of approximation, whereas subsequent ones will
n
u1 c1i 1i + 10 (5.76-a)
i=1
n
u2 c2i 2i + 20 (5.76-b)
i=1
n
u3 c3i 3i + 30 (5.76-c)
i=1
where cji denote undetermined parameters, and are appropriate functions of positions.
46 should satisfy three conditions
1. Be continous.
2. Must be admissible, i.e. satisfy the essential boundary conditions (the natural boundary
conditions are included already in the variational statement. However, if also satisfy
them, then better results are achieved).
3. Must be independent and complete (which means that the exact displacement and their
derivatives that appear in can be arbitrary matched if enough terms are used. Further-
more, lowest order terms must also be included).
In general is a polynomial or trigonometric function.
47We determine the parameters cji by requiring that the principle of virtual work for arbitrary
variations cji . or
n
(u1 , u2 , u3 ) = c1 + 2 c2i + 3 c3i =0 (5.77)
i=1
c1i i ci ci
for arbitrary and independent variations of c1i , c2i , and c3i , thus it follows that
=0 i = 1, 2, , n; j = 1, 2, 3 (5.78)
cji
Thus we obtain a total of 3n linearly independent simultaneous equations. From these displace-
ments, we can then determine strains and stresses (or internal forces). Hence we have replaced
a problem with an innite number of d.o.f by one with a nite number.
48 Some general observations
1. cji can either be a set of coecients with no physical meanings, or variables associated
with nodal generalized displacements (such as deection or displacement).
2. If the coordinate functions satisfy the above requirements, then the solution converges
to the exact one if n increases.
4. Since the strains are computed from the approximate displacements, strains and stresses
are generally less accurate than the displacements.
5. The equilibrium equations of the problem are satised only in the energy sense = 0
and not in the dierential equation sense (i.e. in the weak form but not in the strong one).
Therefore the displacements obtained from the approximation generaly do not satisfy the
equations of equilibrium.
6. Since the continuous system is approximated by a nite number of coordinates (or d.o.f.),
then the approximate system is stier than the actual one, and the displacements obtained
from the Ritz method converge to the exact ones from below.
y
515
111
000 111
000
L
Figure 5.3: Uniformly Loaded Simply Supported Beam Analysed by the Rayleigh-Ritz Method
5.4.1 Examples
L L
M2
= U W = dx wv(x)dx (5.82)
o 2EIz 0
2
Recalling that: M
= d v2 , the above simplies to:
EIz dx
L# 2 $
EIz d2 v
= wv(x) dx (5.83)
0 2 dx2
L
EIz
= (2a1 ) a1 wx(L x) dx
2
0 2
EIz 2 L3 L3
= 4a1 L a1 w + a1 w
2 2 3
a wL 3
1
= 2a21 EIz L (5.84)
6
If we now take
VARIATIONAL and RAYLEIGH-RITZ METHODS
= 0, we would obtain:
a1
wL3
4a1 EIz l = 0
6
wL2
a1 = (5.85)
24EIz
L
Having solved the displacement eld in terms of a1 , we now determine vmax at 2:
wL4 x x2
v = 2
24EI L L
z
a1
wL4
= (5.86)
96EIz
4 4
exact = 5 wL =
This is to be compared with the exact value of vmax wL
384 EIz 76.8EIz which constitutes
17% error.
wL2 w
Note: If two terms were retained, then we would have obtained: a1 = 24EI z
& a2 = 24EI z
exact . (Why?)
and vmax would be equal to vmax
where 1 (t), 2 (t), and 3 (t) are the undetermined parameters. Use the Rayleigh-Ritz method
procedure to generate the governing heat transfer equilibrium equations, using the following
functional L 2 L
1
= k dx q B dx |x=0 q0 (5.88)
0 2 x 0
with the essential boundary condition |x=L = i
Solution:
1. Substituting
L L
1 ! " ! "
= k (2 )2 + 42 3 x + 4(3 )2 x2 dx 1 + 2 x + 3 x2 q B dx 1 q0 (5.89)
0 2 0
2. Invoking
= 0; = 0; =0 (5.90)
1 2 3
we obtain L B
0 0 0 1 0 q dx + q0
k 0 L L2 2 = L B
0 xq dx
(5.91)
4 3
L 2 B
0 L2 3L 3
0 x q dx
3. In this problem q0 varies with time, hence heat capacity must be accounted for
517
q B = c (5.92)
t
4. Substituting
1 2 1 3
0 0 0 1 L 2L 3L 1 q0
k 0 L L 2 1 2
+ c 2 L 1 3 1 4
2 3L 4L
2
= 0
(5.93)
4 3 1 3 1 4 1 5
0 L2 3 L 3 3L 4L 5L 3 0
5. The nal equilibrium equation are now obtained by imposingon the last equation the
condition that |x=L = i , i.e.
Chapter 6
INTERPOLATION FUNCTIONS;
NATURAL COORDINATE
SYSTEMS
6.1 Introduction
1 In the Rayleigh Ritz method we solved the variational problem using a functional approxima-
tion for the displacement eld (Chapter ??). This powerful method has its limitation in terms
of the complexity of solvable problems.
2 As stated above, an alternate approach consists in adopting an approximate displacement
eld in terms of the nodal displacements via interpolation functions (or shape functions).
3 For an element (nite or otherwise), we can write an expression for the generalized displace-
n
u= Ni (X)i = N(x){u} (6.1)
i=1
where:
(a) Ni = 1 at ui
(b) Ni = 0 at uj where i = j.
6.2.1 C0
6.2.1.1 Truss element
or:
63
N1 = 1 x
L (6.11)
N2 = Lx
6.2.1.2 Generalization
where [p] corresponds to the polynomial approximation, and {a} is the coecient vector.
10 We next apply the boundary conditions:
u1 0 1 a1
= (6.13)
u2 L 1 a2
{u} [L] {a}
13Note that in some cases L1 is not always possible to obtain, and that in others there may
be considerable algebraic diculties for arbitrary geometries. Hence, we shall introduce later
on Lagrangian and Hermitian interpolation functions.
14 Next we consider a triangular element, Fig. 6.2 with bi-linear displacement eld (in both x
and y):
u = a1 + a2 x + a3 y (6.17)
v = a4 + a5 x + a6 y (6.18)
a1
u = 1 x y a2 (6.19)
a3
[p]
{a}
15 As before, we rst seek the shape functions, and hence we apply the boundary conditions at
16 We then multiply the inverse of [L] in Eq. 6.20 by [p] and obtain:
u = N1 u1 + N2 u2 + N3 u3 (6.21)
where
1
N1 = (x2 y3 xy3 x2 y + x3 y)
x2 y3
1
N2 = (xy3 x3 y) (6.22)
x2 y3
y
N3 =
y3
We observe that each of the three shape functions is equal to 1 at the corresponding node, and
equal to 0 at the other two.
17 The same shape functions can be derived for v:
v = N1 v 1 + N2 v 2 + N3 v 3 (6.23)
u1
v1
u N1 0 N2 0 N3 0 u2
= (6.24)
v 0 N1 0 N2 0 N3
v2
u3
v3
19 The element is refereed to as Constant Strain Triangle (CST) because it has a linear dis-
placement eld, and hence a constant strain.
3. Inversion of [L]
21 By following these operations, we have in eect dened the Lagrangian Interpolation Func-
tions for problems with C 0 interelement continuity (i.e continuity of displacement only).
22 The Lagrangian interpolation denes the coecients ([N] in our case) of a polynomial series
representation of a function in terms of values dened at discrete points (nodes in our case).
For points along a line this would yield:
Q
m+1
Ni = Q
j=1,j=i (xxj )
m+1 (6.25)
j=1,j=i (xi xj )
24 For the axial member, m = 1, x1 = 0, and x2 = L, the above equations will result in:
(x L) x x x
u= u1 + u2 = (1 ) u1 + u2 (6.27)
L L L
L
N1 N2
25 Next we consider a quadrilateral element, Fig. 6.3 with bi-linear displacement eld (in both
x and y).
26 Using the Lagrangian interpolation function of Eq. 6.25, and starting with the u displace-
ment, we perform two interpolations: the rst one along the bottom edge (1-2) and along the
top one (4-3).
27 From Eq. 6.25 with m = 1 we obtain:
x2 x x1 x
u12 = u1 + u2
x2 x1 x1 x2
ax x+a
= u1 + u2 (6.28)
2a 2a
28 Similarly
x2 x x1 x
u43 = u4 + u3
x2 x1 x1 x2
ax x+a
= u4 + u3 (6.29)
2a 2a
29Next, we interpolate in the y direction along 1-4 and 2-3 between u12 and u43 . Again, we
use Eq. 6.25 however this time we replace x by y:
y2 y y1 y
u = u12 + u43 (6.30)
y2 y1 y1 y2
byax byx+a y+bax y+bx+a
= u1 + u2 + u4 + u3
2b 2a 2b 2a 2b 2a 2b 2a
(a x)(b y) (a + x)(b y) (a + x)(b + y) (a x)(b + y)
= u1 + u2 + u3 + u4
4ab
4ab
4ab
4ab
N1 N2 N3 N1
30 One can easily check that at each node i the corresponding Ni is equal to 1, and all others
to zero, and that at any point N1 + N2 + N3 + N4 = 1. Hence, the displacement eld will be
given by:
u1
v 1
u 2
u N1 0 N2 0 N3 0 N4 0 v2
= (6.31)
v 0 N1 0 N2 0 N3 0 N4 u3
v3
u 4
v4
31 By extension to the previous derivation, the shape functions of a solid rectangular trilinear
u1
v1
w1
u2
v2
u N1 0 0 N2 0 0 N3 0 0 N4 0 0
w2
v = 0 N1 0 0 N2 0 0 N3 0 0 N4 0 (6.32)
u3
w 0 0 N1 0 0 N2 0 0 N3 0 0 N4
v3
w3
u4
v4
w4
where
(a x)(b y)(c z)
Ni = (6.33)
8abc
6.2.2 C1
6.2.2.1 Flexural
32 With reference to Fig. 6.5. We have 4 d.o.f.s, {u}41 : and hence will need 4 shape functions,
N1 to N4 , and those will be obtained through 4 boundary conditions. Therefore we need to
assume a polynomial approximation for displacements of degree 3.
v = a1 x3 + a2 x2 + a3 x + a4 (6.34)
dv
= = 3a1 x2 + 2a2 x + a3 (6.35)
dx
1. v = v 1 at x = 0
2. v = v 2 at x = L
dv at x = 0
3. = 1 = d x
4. = 2 = dv at x = L
dx
or:
v1 0 0 0 1 a1
1 0 0 1 0 a2
=
L3 L2 (6.37)
v 2
L 1 a
3
2 3L2 2L 1 0 a4
{u} [L] {a}
=0 =1
69
where = xl .
37 Hence, the shape functions for the exural element are given by:
N1 = (1 + 2 3 3 2 )
N2 = x(1 )2
N3 = (3 2 2 3 )
N4 = x( 2 ) (6.41)
and are shown in Fig 6.6.
38 Table 6.1 illustrates the characteristics of those shape functions
39 For problems involving the rst derivative of the shape function, that is with C 1 interelement
continuity (i.e continuity of rst derivative or slope) such as for exure, Hermitian interpolation
functions rather than Lagrangian ones should be used.
0.8
N1
0.6
N3
N2
N4
0.4
N
0.2
0.0
0.2
0.0 0.2 0.4 0.6 0.8 1.0
(x/L)
40 Hermitian interpolation functions are piecewise cubic functions which satisfy the conditions
of displacement and slope (C 0 , C 1 ) continuities. They are extensively used in CAD as Bezier
curves.
3. Shape functions should be complete, and meet the same requirements as the coecients
of the Rayleigh Ritz method.
41 Natural coordinates are dimensionless and dened with respect to element length/area/volume
rather than the global coordinate system.
42 They are often used in element formulation of simplex elements (where in an n dimensional
L
x1 x2
L2 L1 = =
O 1 L 2 L
P
x1
1
= x + x
1 1 2 2
x =x+L
2 1
L1 L2
1 = and 2 = (6.42)
L L
since L1 + L2 = L, we have
1 + 2 = 1
x = 1 x1 + 2 x2 (6.43)
L =1
3
L =0
Sid
2 L =0 1
Side
e
A 2
1
A2 P 1
A L =1
3
2
Side 3
L =1 L =0
1 3
49 For a triangular surface, any point P divides the triangle 1-2-3 into three subareas A1 , A2 , A3 ,
A1 A21 A3
L1 = , L2 = , L3 = (6.48)
A A A
1
51 The centroid is at L1 = L2 = L3 = 3
inverting
L1 x2 y3 x3 y2 y2 y3 x3 x2 1
L = x3 y1 x1 y3 y3 y1 x1 x3 x (6.51-a)
2
L3 x1 y2 x2 y1 y1 y2 x2 x1 y
[A]1
2A23 y23 x32 1
1
= 2A31 y31 x13 x (6.51-b)
2A
2A12 y12 x21 y
where xij = xi xj
and if the labels are reversed such that 1-2-3 is clockwise, then the determinant is negative.
54 A function may be expressed in terms of area coordinates
= (L1 , L2 , L3 ) (6.53-a)
L1 L2 L3
= + + (6.53-b)
x L1 x L2 x L3 x
L1 L2 L3
= + + (6.53-c)
y L1 y L2 y L3 y
and
L1 y2 y3 L2 y3 y1 L3 y1 y2
x = 2A ; x = 2A ; x = 2A ;
L1 x3 x2 L2 x1 x3 L3 x2 x1 (6.54)
y = 2A ; y = 2A ; y = 2A ;
55 Integration for the stiness matrix will require integration in area coordinate of expressions
of the form A Lk1 Ll2 Lm3 dA. To accomplish this operation, and with reference to Fig. 6.9,
(Eisenberg and Malvern 1973):
h2
3 L =0
1
11
00
00
11
00
11
L =0 00
11
2 00
11
00
11
s 00
11 s
2 1
ds 1
2
ds
2
1 L =0
3
h1
1 1L1
Thus
Lk1 Ll2 Lm
3 dA = 2A Lk1 Ll2 (1 L1 L2 ) dL2 dL1
m
(6.56)
A 0 0
each of the integrals on the right-hand side is of the form of the beta function (Abramowitz
and Stegun 1970)
1
(z)(w)
B(z, w) = tz1 (1 t)w1 dt = (6.58)
0 (z + w)
where denotes the gamma function which satises (n + 1) = n! for integers n 0. Thus
(k + 1)(l + 1)(m + 1)
Lk1 Ll2 Lm
3 dA = 2A (6.59)
A (k + l + m + 3)
56 It should be noted that the above relation is used only if we have straight sided elements and
the element formulation can thus be analytically derived. Alternatively, if we have a curvilinear
side, this formula will not be used, and the element will be numerically integrated within the
context of an isoparametric formulation.
61 Triangular elements allow a complete polynomial in Cartesian coordinate to be used for the
eld (e.g. temperature, displacement) quantity. Thus, all terms of a truncated Pascal triangle
are used in the shape functions. (This will not be the case for the bilinear quadrilateral element).
)
62 We seek to determine = Ni i where i are nodal d.o.f. and Ni = Ni (L1 , L2 , L3 ).
)n q r s
63 We start with =
i=1 ai L1 , L2 , L3 ) where q, r and s range over the n possible combinations
for which q + r + s = p and where p is the order of the polynomial. For example for
Linear element:
= a1 L1 + a2 L2 + a3 L3 = a 1 + a 2 x + a 3 y (6.63)
Quadratic element:
64 The shape functions for the Linear Triangle are simply the area coordinates,
N1 = L1 N2 = L2 N3 = L3 (6.65)
and each one is equal to unity at one node, zero at the others, and varies linearly.
65 The shape functions for other elements can be obtained from the Lagrangian interpolation
66 Hence, denoting a typical node i by three numbers q, r and s corresponding to the position
of coordinates L1i , L2i and L3i , we can write the shape functions in terms of three Lagrangian
interpolation functions
Ni = lqq (L1 )lrr (L2 )lss (L3 ) (6.67)
67 Using this formula, the shape functions for higher order elements are
Linear triangle
Ni = Li (6.68)
Quadratic triangle
Cubic triangle
Constant a1
Linear a2 x a3 y
Quadratic a4 x2 a5 xy a6 y 2 (6.71)
3 2 2
Cubic a7 x a8 x y a9 xy a10 y 3
Quartic a11 x4 a12 x3 y a13 x2 y 2 a14 xy 3 a15 x4
Table 6.2: Interpretation of Shape Functions in Terms of Polynomial Series (1D & 2D)
Table 6.3: Polynomial Terms in Various Element Formulations (1D & 2D)
70 A complete polynomial contains all the terms above a certain line in the Pascal triangle.
71 The more terms are included, the higher the accuracy. For instance u = a1 +a2 x+a3 y +a5 xy
is more accurate than u = a1 +a2 x+a3 y, however the rate of convergence is unchanged (because
the second approximation is not a complete quadratic one).
72 Terms in the approximation which do not improve the rate of convergence are called parasitic
terms.
Chapter 7
FINITE ELEMENT
DISCRETIZATION and
REQUIREMENTS
7.1 Discretization
This section is mostly extracted from (Reich 1993)
7.1.1 Discretization of the Variational Statement for the General TPE Vari-
ational Principle
1 The discretization of Equation 5.8 will be performed on an element domain e using the
procedures described in Chapter 2 of (Zienkiewicz and Taylor 1989);
2 The surface of the element subjected to surface tractions t comprises one or more surfaces
of the element boundary . For the present time this discussion will be kept on a very general
level with no mention of the dimensionality of the elements; the number of nodes dening the
elements; or the nature of the constitutive law.
3 The rst step in the discretization process is to dene the displacements u at a point inside
the element in terms of the shape functions N and the nodal displacements ue for the element
u = Nue (7.1)
4 The virtual displacements u at a point inside the element can also be dened in terms of
the shape functions N and the nodal virtual displacements ue for the element
u = Nue (7.2)
dening the virtual strain energy for the element due to the nodal displacements u, the strains
at a point inside the element are expressed in terms of the nodal displacements ue using
Equation 7.1
= Lu = LNue (7.4)
and the virtual strains at a point inside the element are expressed in terms of the nodal
virtual displacements ue using Equation 7.2
def
B = LN (7.6)
8 In order to discretize the volume integrals in Equation 7.3 dening the virtual strain energy
for the element due to the initial strains 0 and stresses 0 , Equations 7.5 and 7.6, which dene
the virtual strains at a point inside the element in terms of the nodal virtual displacements
ue , are substituted into the integrands
T T
(Lu) D 0 d = ue BT D 0 d (7.10)
e e
(Lu)T 0 d = uTe BT 0 d (7.11)
e e
10 In order to discretize the volume integral dening the work done by the body forces and the
73
surface integral dening the work done by the surface tractions in Equation 7.3, Equation 7.2
is substituted into the integrands
uT bd = uTe NT bd (7.14)
e e
uT td = uTe NT td (7.15)
t t
the sum of the internal and external virtual work due to body forces and surface tractions is
T
u bd + uT td = uTe fe (7.17)
e t
12 Having obtained the discretization of the various integrals dening the variational statement
for the TPE variational principle, it is now possible to dene the discrete system of equa-
tions. Substituting Equations 7.7, 7.13, and 7.17 into Equation 7.3 and rearranging terms, the
discretized Principle of Virtual Work is
13Since uTe is an arbitrary (i.e. non-zero) vector appearing on both sides of Equation 7.18,
the discrete system of equations can be simplied into
Ke ue = fe + f0e + Pu (7.19)
14 The discretization of the three variational statements dened in Equation 5.65, 5.66, and
5.67 will be performed on an element domain e using the procedures described in Chapter
2 of (Zienkiewicz and Taylor 1989) assembly of the discrete element equations into a discrete
global system of equations is straightforward and will be omitted from this discussion.
15 The surface of the element subjected to surface tractions t comprises one or more surfaces
of the element boundary e . For the present time this discussion will be kept on a very general
level with no mention of the dimensionality of the elements; the number of nodes dening the
elements; or the nature of constitutive law.
16 The rst step in the discretization process is to dene the displacements u, strains , and
stresses at a point inside the element in terms of the shape functions Nu , N= , and N ,
respectively, and the element nodal displacements ue , strains e , and stresses e
u = Nu ue (7.20)
= N = e (7.21)
= N e (7.22)
We note that contrarily to the previous case (Eq. 7.1) we now have three discretizations (instead
of just one).
17 The virtual displacements u, virtual strains , and virtual stresses at a point inside
the element can also be dened in terms of the shape functions Nu , N= , and N , respectively,
and the nodal virtual displacements ue , virtual strains e , and virtual stresses e for the
element
u = Nu ue (7.23-a)
= N= e (7.23-b)
= N e (7.23-c)
19 In order to discretize the volume integral in the rst variational statement (i.e. Equ. 7.24-a)
dening the virtual strain energy for the element, Equation 7.23-a is substituted into the virtual
strain-displacement relationship (i.e. Equation 5.6) to dene the virtual strains at a point
inside the element in terms of the nodal virtual displacements ue
(L u) = L u = L Nu ue (7.25)
Bu = L Nu (7.26)
and substituting Equation 7.22 into the integrand, the virtual strain energy for an element is
written as
(Lu) d = ue
T T
BTu N d e (7.27)
e e
22 In order to discretize the volume integral dening the work done by the body forces and the
75
surface integral dening the work done by the surface tractions in the rst variational statement
(i.e. the rst equation in Equation ??), Equation 7.23-a is substituted into the integrands
T T
u b d = ue NTu b d (7.30)
e e
uT t d = uTe NTu t d (7.31)
t t
23 Having dened the discretization of the various integrals in the rst variational statement
for the HW variational principle (i.e. Equ. 7.24-a), it is now possible to dene the discrete
system of equations. Substituting Equations 7.29 and 7.33 into the variational statement and
rearranging terms, the discretized Principle of Virtual Work is
where the left-hand side is the virtual strain energy and the right-hand side is the internal and
external virtual work. Since ue is an arbitrary (i.e. non-zero) vector appearing on both sides
of Equation 7.34, the discrete system of equations can be simplied into
FTe e = fe (7.35)
Since the nodal virtual strains are arbitrary they can be eliminated from Equation 7.40
yielding
Ae e Ce e = ge (7.41)
Recognizing that
NT Bu d = Fe (7.43)
e
NT N= d = CTe (7.44)
e
Since the nodal virtual stresses e are arbitrary they can be eliminated from Equation 7.45
yielding
Fe ue CTe e = 0 (7.46)
dene the discrete mixed system of equations for an element. Assembling Equations 7.35, 7.41,
and 7.46 in matrix form adopting the classic arrangement for a constrained system of equations
Ae Ce 0 e ge
CTe 0 Fe e = 0 (7.47)
T
0 Fe 0 ue fe
28 Since it would be computationally expensive to solve the system of equations in Eq. 7.47
77
using direct method, an indirect or iterative procedure (i.e. Gauss-Seidel instead of Gauss-
Jordan) is often selected, (Zienkiewicz and Taylor 1989).
Step 1: uk+1
n = ukn + K1 rkn
Step 2: k+1
n = CT Fuk+1
n
(7.48)
Step 3: k+1
n = C1 Ak+1
n
Step 4: rk+1
n = f FT k+1
n
31A nite element (just a an approximate displacement eld in the Rayleigh-Ritz formulation)
must satisfy two basic requirements
Completeness: The FE discretization must at least accommodate constant displacement and
constant strain (or temperature and temperature gradient). This is accomplished by
including in two dimensional problems the following
Compatibility or Conformity: The approximation of the eld over element boundaries must
be continuous (C0 or C1 continuity). Most nite elements are conforming, but some are
not.
32 For instance, with respect to Fig. 7.1, element A must be capable of undergoing rigid
body motion without internal strains/stresses, and at node B we should have continuity of
displacement (but not slope for this element).
33 If those two requirements are satised, then convergence is assured. In the FE method
approximate solution are obtained, and the more elements we use, the more accurate is the
111
000
000
111
000
111
000
111
000
111
P
111
000
000
111
000
111
000
111
000
111 B
approximate solution. In the limit, for innitely small elements, we require the solution to
be also innitely close to the exact one. Hence, convergence is ensured if completeness and
compatibility requirements are satised.
34 Lax equivalence theorem:
Convergence criterion=completeness+ compatibility requirements
Patch test: Completeness can be assessed through the patch test which will be discussed
later.
Zero Strain Energy: For structural problems, there should be zero strain energy when the
element is subjected to a rigid body motion. Recall that in the stiness matrix formu-
lation, the matrix is singular as it embodies not only the force displacement relations,
but the equilibrium equations also. To each of those equations, corresponds a rigid body
mode which can be detected by an eigenvalue analysis (more about this later).
36 Approximation will yield an exact solution in the limit as the size h of element approaches
zero.
37 In some cases the exact solution is obtained with a nite number of elements (or even with
only one) if the polynomial expansion used in that element can t exactly the correct solution.
(e.g. Truss, beam elements, plane stress element used in a plate under pure axial load).
38 The exact solution can always be expanded in the vicinity of any node i as a polynomial
u u
u = ui + (x xi ) + (y yi ) + ... (7.51)
x i y i
If within an element of size h a polynomial of degree p is used, then the Taylor expansion up
to that degree can be accommodated and the error is of the order O(hp+1 ).
39
79
Thus for a linear element, the convergence rate is of order O(h2 ), i.e., the error in displace-
ment is reduced to 1/4 of the original error by halving the mesh.
40 By a similar argument, it can be shown that the strains (or stresses) which are given by
the mth derivatives of the displacement should converge with an error of O(hp+1m ), and the
strain energy (which is given by the square of the stresses) error will be O(h2(p+1m) )
41 From the above arguments, knowledge of the order of convergence may help in extrapolating
the solutions to the correct one. Hence, for instance, if the displacement converges at O(h2 ),
and we have two approximate solutions u1 and u2 obtained with meshes of sizes h and h/2,
then we can write
u1 u O(h2 )
= =4 (7.52)
u2 u O(h/2)2
where u is the extrapolated value.
42 Convergence to the exact solution can be accelerated by, Fig. 7.2.
p
h
p increasing the order of the polynomial (same number of elements but higher number of
nodes/dof).
A third form of acceleration is the so called r renement in which the same number of nodes/elements
is retained but the mesh is shifted around to increase its density in zones of high stress gradient.
43 The above convergence procedures can be accelerated within the context of a program which
7.4
FINITE ELEMENT DISCRETIZATION and REQUIREMENTS
45 A numerical solution that is derived from the principle of minimum total potential energy is
a lower bound solution, because the strain energy is smaller than the exact one (i.e. obtained
from an innite number of elements).
46 This can be readily shown if we consider the displacement ui caused by a load Pi which is
increased from zero to its stipulated value. The work done is Pi2ui and must be equal to the
internal strain energy U . Alternatively, the potential of the applied load is Pi ui , and the exact
potential energy is
Pi uexact Pi uexact
exact = i
Pi uexact
i = i
(7.53)
2 2
similarly, the approximate value of the potential energy is
Pi uapprox
approx = i
(7.54)
2
We know that the approximation of is algebraically higher than the exact value (since the
exact value is a minimum), hence
Pi uiexact Pi uiapprox
< (7.55)
2 2
exact approx
or
uapprox
i < uexact
i (7.56)
48 In an exact solution over each dierential element equilibrium and compatibility prevails.
49 In a nite element discretization, this is not necessarily the case:
Compatibility at Nodes: Elements connected to one another have the same displacements
(along corresponding dof) at the connecting node.
Interelement Continuity of Stress: is usually not satised, that is equilibrium across the
elements does not necessarily prevail. For example consider Fig. 7.3 if node 4 is the only
one displaced (all others remaining xed), then there will be a discontinuity of xx across
element boundary 2-3.
3
711
x2
1 1 2 4
u
4
2
4 v
2
y,v
x,u
1 v
3
5
Chapter 8
8.1 Introduction
1 Having rst introduced the method of virtual displacements, than the interpolation (or shape)
functions [N], which relate internal to external nodal displacements, and nally having applied
the virtual displacement method to nite element in chapter 7, we derive the stiness matrices
of some simple elements.
2 The shape functions of the truss element were derived in Eq. 6.11:
x
N1 = 1
L
x
N2 =
L
du
xx =
dx
= [ dN1 dN2 ]
dx dx
= [ L1 L1 ] (8.1)
[B]
L
L1
e
[k ] = A 1 E L1 1
L dx (8.2-a)
0 L
L
AE 1 1
= dx
L2 0 1 1
AE 1 1
= (8.2-b)
L 1 1
5 We observe that this stiness matrix is identical to the one earlier derived in Eq. ??.
6 For a beam element, for which we have previously derived the shape functions in Eq. 6.41
8 Using the shape function for the beam element from Eq. 6.41, and noting the change of
integration variable from dx to d, we obtain
6
L2 (2 1)
1
. /
L2 (3 2)
[ke ] = EIz 6
6
L2 (2 1) L2 (3 2) 6
L2 (2 + 1) L2 (3 1) Ld (8.6)
0
L2 (2 + 1)
L2 (3 1) dx
or
v1 1 v2 2
V1 12EI z 6EIz
12EI z 6EIz
6EIL3 L2 L3 L2
M z 4EIz
L2
6EIz 2EIz
(8.7)
[ke ] =
1 L2 L L
V2 12EIz
6EI L3
6EI
L2
z 12EIz
L3
6EI
L2
z
M2 2
z
L
2EIz
L 6EI
L2
z 4EIz
L
10 Having retrieved the stiness matrices of simple one dimensional elements using the principle
of virtual displacement, we next consider two dimensional continuum elements starting with
the triangular element of constant thickness t made out of isotropic linear elastic material. The
element will have two d.o.fs at each node:
{u} = u1 u2 u3 v1 v2 v3 T (8.8)
{} = xx yy xy T (8.9)
hence we can rewrite the strains in terms of the derivatives of the shape functions through the
matrix [B]:
N
xx 0
x N u
= 0 y (8.10)
yy N N v
xy y x
[B]
13 We note that because we have 3 u and 3 v displacements, the size of [B] and [u] are 9(33)6
and 6 1 respectively.
14
STRAIGHT SIDED ELEMENTS; 1st GENERATION
15 With the constitutive matrix [D] given by Eq. ??, the strain-displacement relation [B] by
Eq. 8.11, we can substitute those two quantities into the general equation for stiness matrix,
Eq. 7.8:
Z
[ke ] = [B]T [D][B]d
2 x3 x2 3
x12 0 x2 y 3
6 1
xx 3 7 2 32 3
Z 6 x2
0 2 y3 7 x12 1
6 1 7 E 1 0 x2
0 0 0 0
6 0 0 y3 7 4 1 56 x3 x2
xx 7 tdxdy
= 6 x3 x2 7 0 4 0 0 0 3 1
5
6 7 1 2 | {z }
x2 y 3 2 y3 y3
0 x12
6 x2 y 3 7 0 0 1
) x3 x2
xx 3 1
x12 1
0
4 0 xx 3 1 5| {z 2
}| x2 y 3 2 y3 y3
{z
x2
} dV ol
2 y3 x2
1 [D]
0 y3
0 [B]
| {z }
[B]T
2 3
y32 + x232 y32 x3 x32 x2 x32 y3 x32 x3 y3 + y3 x32 x2 y3
6 y32 x3 x32 y32 + x23 x2 x3 y3 x32 + x3 y3 x3 y3 x2 y3 7
6 7
6 x2 x32 x2 x3 x22 x2 y3 x2 y3 0 7
= 6
6
7
7
6 y3 x32 y3 x32 + x3 y3 x2 y3 y32 + x232 y32 x3 x32 x2 x32 7
4 x3 y3 + y3 x32 x3 y3 x2 y3 y32 x3 x32 y32 + x23 x2 x3 5
x2 y3 x2 y3 0 x2 x32 x2 x3 x22
where = 1
2 , = 1+
2 , = ET
2(1 2 )x2 y3
, x32 = x3 x2 , and y32 = y3 y2 .
16 For the evaluation of the internal stresses {} = [D][B]{u} hence for this particular element
we will have:
u1
x2 1 1
u2
x 1 0 x2 0 0 0 0
E
0 x x
x 1 u3
0 0 y3
3 2 3
y = 1 0
1 2 x3 x2 x3
x 2 y3 x 2 y3
v1
xy 0 0 1 ) 1
1 1
0
2
x 2 y3 x 2 y3 y3
x2 x2
v2
{ } [D] [B] v3
{u}
u1
u2
y3 y3 0 x32 x3 x2
u
= y3 y3 0 x32 x3 x2 3
(8.12)
v1
x32 x3 x2 y3 y3 0
v
2
Victor Saouma v3 II; Solid Mechanics
Finite Elements
Draft
8.3 Triangular Elements
where = E
85
(1 2 )x2 y3
17From Eq. 8.12 we note that the element must be positioned in such a way that neither x2
nor y3 is equal to zero.
18 A palliative to this problem is to use a formulation based on natural coordinate systems.
8.3.2.1 Linear, T3
3
u = Li ui = Ni ui
i=1 i=1
(8.13)
3
v = Li v i = Ni v i
i=1 i=1
or Ni = Li .
21 The displacement shape functions are given by Equation 6.51-b:
L1 1
2A23 y23 x32
L2 = 2A31 y31 x13 (8.14)
2A
L3 2A12 y12 x21
where xij = xi xj .
22 Derivatives required for strain-displacement relationships are obtained from Eq. 6.54:
L1 y2 y3 L2 y3 y1 L3 y1 y2
x = 2A ; x = 2A ; x = 2A ;
L1 x3 x2 L2 x1 x3 L3 x2 x1 (8.15)
y = 2A ; y = 2A ; y = 2A ;
25 This expression is then analytically integrated using Eq. 6.60. In general an isoparametric
8.3.2.2
STRAIGHT SIDED ELEMENTS; 1st GENERATION
6
u = Ni ui v = Ni vi (8.18)
i=1 i=1
27 It should be noted that alternatively, we could have obtained those shape functions by
assuming
u(L1 , L2 ) = a1 + a2 L1 + a3 L2 + a4 L21 + a5 L1 L2 + a6 L22 (8.20)
and then determine the coecients through satisfaction of the appropriate boundary conditions.
28 The shape functions and their derivatives with respect to the natural coordinates L1 and L2
are given in Table 8.1 (in terms of L1 , L2 , and L3 ).
or
xx xx (L1 , L2 , L3 )
= (L , L , L ) (8.22-a)
yy yy 1 2 3
xy xy (L1 , L2 , L3 )
6
x
Ni (L1 ,L2 ,L3 ) ui
= 0 y (8.22-b)
Ni (L1 ,L2 ,L3 ) Ni (L1 ,L2 ,L3 )
vi
i=1
y x
B=LN
L1
x
L2
x
L3
x
[NL ] 0
1
[NL ]
u
= 0 L1
y
L2
y
L3
y (8.22-c)
2A L1 L2 L3 v
y y y [NL ] L1
x
L2
x
L3
x [NL ]
where
N1 N2 N3 N4 N5 N6
L1 L1 L1 L1 L1 L1
N1 N2 N3 N4 N5 N6
[NL ] = L2 L2 L2 L2 L2 L2 (8.23-a)
N1 N2 N3 N4 N5 N6
L3 L3 L3 L3 L3 L3
4L1 1 0 0 4L2 0 4L3
= 0 4L2 1 0 4L1 4L3 0 (8.23-b)
0 0 4L3 1 0 4L2 4L1
where
x (4L1 1) y (4L1 1)
L1 L1
x (4L2 1) y (4L2 1)
L2 L2
! " 1
x (4L3 1)
L3
y (4L3 1)
L3
NTx |NTy = (8.25-a)
2A
4 L 1 L2
x L2 + 4 x L1 4 L 1 L2
y L2 + 4 y L1
4 L L3
x L3 + 4 x L2
2
4 L L3
y L3 + 4 y L2
2
4 x L1 + 4 L
L3
x L3
1
4 y L1 + 4 L
L3
y L3
1
u v = u1 u2 u3 u4 u5 u6 v 1 v 2 v 3 v 4 v 5 v 6 (8.25-b)
31 This expression is then analytically integrated using Eq. 6.60. In general an isoparametric
32The structure stiness matrix of the bilinear rectangular shown in Fig. 8.1 is described by
the Mathematica code shown below.
N3[x_,y_,a_,b_] := (1/4)(1+(x/a))(1+(y/b));
N4[x_,y_,a_,b_] := (1/4)(1-(x/a))(1+(y/b));
Visualize the shape functions;
Plot3D[N1[x,y,1,2], {x, -1, 1}, {y, -2, 2}];
Plot3D[N3[x,y,1,2], {x, -1, 1}, {y, -2, 2}];
Define the differential operators;
diffx[f_] := D[f,x];
diffy[f_] := D[f,y];
Shape Function Matrix
Nmat = {
{N1[x,y,a,b],0,N2[x,y,a,b],0,N3[x,y,a,b],0,N4[x,y,a,b],0},
{0,N1[x,y,a,b],0,N2[x,y,a,b],0,N3[x,y,a,b],0,N4[x,y,a,b]}
}
B Matrix
Bmat = {Map[diffx,Nmat[[1]]],Map[diffy,Nmat[[2]]],
Map[diffy,Nmat[[1]]]+Map[diffx,Nmat[[2]]]}
Define the constitutive matrix
Dmat = E/(1-nu^2) *
{{1,nu,0},{nu,1,0},{0,0,(1-nu)/2}}
Determine the stiffness matrix
Intmat = Transpose[Bmat] . Dmat . Bmat;
Kmat =Simplify[ t * Integrate[ Integrate[Intmat, {y, -b, b}], {x, -a, a}] ]
TeXForm[Kmat]
ke18;14 =
6 a b (1+) (1+) 8 (1+) 12 a b (1+) (1+) 8 (1+) (1+)
(e t) (e (2 a2 +b2 b2 ) t) e (1+3 ) t (e (a2 b2 +b2 ) t)
8 (1+) 6 a b (1+) (1+) 8 (1+) (1+) 6 a b (1+) (1+)
e (a2 +4 b2 +a2 ) t e (1+3 ) t e (a2 2 b2 +a2 ) t
et
12 a b (1+) (1+) 8 (1+) (1+) 6 a b (1+) (1+) 8 (1+)
(e (1+3 ) t) (e (a2 b2 +b2 ) t) (e (2 a2 +b2 b2 ) t)
et
8 (1+) (1+) 6 a b (1+) (1+) 8 (1+) 6 a b (1+) (1+)
(e (a2 2 b2 +a2 ) t) (e (a2 +b2 +a2 ) t)
et e (1+3 ) t
12 a b (1+) (1+) 8 (1+) 6 a b (1+) (1+) 8 (1+) (1+)
e (2 a2 +b2 b2 ) t (e (1+3 ) t) e (4 a2 b2 +b2 ) t
et
8 (1+) 12 a b (1+) (1+) 8 (1+) (1+) 12 a b (1+) (1+)
(e (a2 +b2 +a2 ) t) (e (1+3 ) t) (e (a2 2 b2 +a2 ) t) (e t)
6 a b (1+) (1+) 8 (1+) (1+) 12 a b (1+) (1+) 8 (1+)
e (1+3 ) t e (4 a2 b2 +b2 ) t (e t) e (2 a2 +b2 b2 ) t
8 (1+) (1+) 12 a b (1+) (1+) 8 (1+) 12 a b (1+) (1+)
ke18;58 =
(e (a2 2 b2 +a2 ) t) (e (a2 +b2 +a2 ) t)
et e (1+3 ) t
12 a b (1+) (1+) 8 (1+) 6 a b (1+) (1+) 8 (1+) (1+)
e (2 a2 +b2 b2 ) t (e (1+3 ) t) e (4 a2 b2 +b2 ) t
et
8 (1+) 12 a b (1+) (1+) 8 (1+) (1+) 12 a b (1+) (1+)
(e (a2 +b2 +a2 ) t) (e (1+3 ) t) (e (a2 2 b2 +a2 ) t) (e t)
6 a b (1+) (1+) 8 (1+) (1+) 12 a b (1+) (1+) 8 (1+)
e (4 a2 b2 +b2 ) t (e t) e (2 a2 +b2 b2 ) t
e (1+3 ) t
8 (1+) (1+) 12 a b (1+) (1+) 8 (1+) 12 a b (1+) (1+)
e (a2 2 b2 +a2 ) t (e t) e (a2 +4 b2 +a2 ) t (e (1+3 ) t)
6 a b (1+) (1+) 8 (1+) 12 a b (1+) (1+) 8 (1+) (1+)
(e t) (e (2 a2 +b2 b2 ) t) (e (a2 b2 +b2 ) t)
e (1+3 ) t
8 (1+) 6 a b (1+) (1+) 8 (1+) (1+) 6 a b (1+) (1+)
e (a2 +4 b2 +a2 ) t e (a2 2 b2 +a2 ) t
e (1+3 ) t et
12 a b (1+) (1+) 8 (1+) (1+) 6 a b (1+) (1+) 8 (1+)
(e (1+3 ) t) (e (a2 b2 +b2 ) t) et (e (2 a2 +b2 b2 ) t)
8 (1+) (1+) 6 a b (1+) (1+) 8 (1+) 6 a b (1+) (1+)
(8.26-a)
34 The CST gives good results in a region of the FE model where there is little strain gradient.
displacement relation was assumed resulting in a constant strain and stress (for linear elastic
material).
37 If we consider the element shown in Fig. 8.2 we apply a moment M such that: 1) origin does
not move, and 2) the displacement is u = u at x = l on top of the beam, then from the exact
solution (Timoshenko and Goodier 1970), the correct displacements are
u u
u= xy v= (x2 y 2 ) (8.27)
cl 2cl
similar descriptions for v, and determining the strains from = Bu for the CST element, we
obtain
1 c
xx = 0 yy = 0 xy = 2 u (8.29)
c 4l
which are clearly wrong. Hence, we conclude the CST element can not properly perform in
bending problems (As is expected from a constant strain element behavior when used in a linear
strain problem).
39 This deciency will later be addressed by including an additional internal drilling degree
of freedom.
40 Zero energy mode can be veried by considering for example the strain xx which is given by
1
xx = (y3 u1 + y3 u2 ) (8.30)
x2 y3
since u1 = u2 for a rigid body motion, we see that xx = 0, similar conditions can be found for
yy and xy
41 Under exure, the top and bottom sides remain straight giving rise to erroneous strains.
42 From Pascalss triangle (Sect. 6.4), the displacement eld for a bilinear rectangular element
is given by
u = a1 + a2 x + a3 y + a5 xy (8.31)
which is consistent with the element shape functions derived earlier (Eq. 6.31)
43 We note that whereas there is a quadratic term (a5 xy) convergence is governed by the full
linear expansion, and that this last term may cause parasitic eects.
formulations.
48Renement of this element include the addition of two degrees of freedom to allow a state of
constant curvature
49 Other renements (mostly in shells) include the inclusion of additional rotational degrees of
freedom termed drilling dof.
Chapter 9
ISOPARAMETRIC ELEMENTS;
2nd GENERATION
9.1 Introduction
1 We have previously examined simple nite elements, in this lecture we shall distort those
(-1,-1) (1,-1)
L 2 =0 L 1 =0
1 2
L 3 =0
y,v
ISOPARAMETRIC ELEMENTS; 2nd GENERATION
x,u Mapping (-1,1) (1,1)
(-1,-1) (1,-1)
u = u v w T = [N]ue (9.2)
6 When elements are also distorted, the coordinates of any point can also be expressed in terms
of nodal coordinates
x 1
x = N1 x1 + N2 x2 + = N x 2 = Nx
..
.
y1
y = N1 y 1 + N2 y 2 + = N y2 = Ny (9.3)
..
.
z1
z = N1 z 1 + N2 z 2 + = N z2 = Nz
..
.
or
c = x y z T = [N]c (9.4)
7 Shape function matrices [N] and [N] are functions of , , and . Hence, three possible cases
Geometric Displacement
Nodal Nodal
Config. Config.
Isoparametric
Subparametric
Superparametric
9 The simplest introduction to isoparamteric elements is through a straight three noded quadratic
1 3 2
x,u
L
10 The shape functions for the element can be obtained from the Lagrangian interpolation
function used earlier, Eq. 6.25, and in which we substitute x by . The kth term in a polynomial
of order n 1 would be
ISOPARAMETRIC ELEMENTS; 2nd GENERATION
-n
( i )
Nkn = -ni=1,i
=k (9.5-a)
i=1,i
=k (k i )
( 1 )( 2 ) ( k1 )( k+1 ) ( n )
= (9.5-b)
(k 1 )(k 2 ) (k k1 )(k k+1 ) (k n )
For a three noded quadratic element 1 = 1, 2 = +1, and 3 = 0. Substituting, we obtain
the three shape functions
(2 )(3 ) (1)(0)
N1 () = (1 2 )(1 3 ) = (11)(10) = 1 2
2 ( )
(1 )(3 ) (+1)(0) 1 2
N2 () = (2 1 )(2 3 ) = (1+1)(10) = 2 ( + ) (9.6)
(1 )(2 ) (+1)(1)
N3 () = (3 1 )(3 2 ) = (0+1)(01) = 1 2
Hence,
x() = N x1 x2 x3 T and u() = N u1 u2 u3 T (9.7)
where
N = 1 2
2 ( ) 1 2
2 ( + ) 1 2 (9.8)
11 The strain displacement relation is given by Eq. 7.4, = Lu = LNue = Bue , and the
dierential operator L is equal to d . For this one dimensional case, this reduces to
dx
du d u1
x = = N u (9.9)
dx dx 2
u3
L
B
12 We invoke the chain rule since the shape functions are expressed in terms of natural coordi-
nates:
dN dN d
B= = (9.10)
dx d dx
The rst term may be readily available from the shape functions, Eq. 9.8, however the second
one is not.
13 Whereas,
d is not available, we may determine its inverse dx , from Eq. 9.7, which we shall
dx d
denote by J or Jacobian. The Jacobian operator J is a scale factor which relates cartesian to
natural coordinates dx = Jd.
dx d x1 x1
J() = = N x2 = 1 (2 1) 12 (2 + 1) 2 x2 (9.11)
d d 2
x3 x3
dN
d
dx
d
1 d 1
B() = N = 1
2 (2 1) 1
2 (2 + 1) 2 (9.13)
J d J
16 Substituting, the element stiness matrix is nally obtained from Eq. 7.8
L +1
T
Ke () = B ()AEB()dx = BT ()AEB()J()d (9.15)
0 1
17 We observe that B, in general, contains terms in both the numerator and denominator, and
hence the expression can not be analytically inverted. Furthermore, the limits of integration
are now from -1 to +1, and we shall see later on how to numerically integrate it.
18 A simple Mathematica code to generate the stiness matrix of three noded (quadratic)
element:
x1=0;x2=L;x3=L/2;
N1[x_,L_]:=(x-x2) * (x-x3)/( (x1-x2) * (x1-x3));
N2[x_,L_]:=(x-x1) * (x-x3)/( (x2-x1) * (x2-x3));
N3[x_,L_]:=(x-x2) * (x-x1)/( (x3-x2) * (x3-x1));
BMat={D[N1[x,L],x], D[N2[x,L],x], D[N3[x,L],x]};
Integrate[EA Outer[Times, BMat, BMat], {x,0,L}];
MatrixForm[%]
9.2.2 Quadrilaterals
9.2.2.1 Linear Element (Q4)
19 We have previously derived the stiness matrix of a rectangular element (aligned with the
= 1 1 2
1 2
y,v
x,u
m
u(, ) = Nij uk = Nj ()Ni ()uk (9.16)
i=1 j=1
N1 (, ) = 1
4 (1 )(1 ); N2 (, ) = 1
4 (1 + )(1 );
(9.19)
N3 (, ) = 1
4 (1 + )(1 + ); N4 (, ) = 1
4 (1 )(1 + );
It should be noted that for this simple case, the shape functions could have been determined
by mere inspction.
23 Coordinates and displacements are given by
) )
x = ) Ni (, )xi ; y = ) Ni (, )y i
(9.20)
u = Ni (, )ui ; v = Ni (, )v i
= LNu = Bu
Ni (,)
xx (, )
4 x 0
Ni (,) ui
yy (, ) = 0 (9.23-a)
y vi
xy (, ) i=1 Ni (,)
y
Ni (,)
x
u
B=LN
u1
v1
u2
N1,x 0 N2,x 0 N3,x 0 N4,x 0
v2
= 0 N1,y 0 N2,y 0 N3,y 0 N4,y
u3
N1,y N1,x N2,y N2,x N3,y N3,x N4,y N4,x
v3
B
u4
v4
Ni (,)
0 # $
4
Ni (,)
= 0 x
x
(9.23-b)
i=1 Ni (,) Ni (,) x y
[J]1
26 Considering the local set of coordinates , and the corresponding global one x, y, the chain
This last equation is the key to get all the components which will go inside the B matrix.
27 Expanding the denition of the Jacobian
N (,) # $ # $
i x y Ni
4 Ni Ni Ni
x xi y i x
Ni (,) = x y Ni = Ni Ni Ni (9.25-a)
y xi y i y
i=1
J
# $ x1 y1
N1 N2 N3 N4 x2 y2 Ni
= x (9.25-b)
N1 N2 N3 N4 x3 y3 Ni
y
x4 y4
x1 y1
1 (1 ) (1 ) (1 + ) (1 + ) x2 y2 Ni
= x
4 (1 ) (1 + ) (1 + ) (1 ) x3 y3 Ni
y
x4 y4
J
(9.25-c)
4
1 y
y Ni
y i N i
y i
[J]1 =
def x x
= = (9.26)
y y detJ x
x
detJ N i
xi
Ni
xi
i=1
29 From calculus, if and are some arbitrary curvilinear coordinates, Fig. 9.6, then
y
B
y
d C
ds
de
y
A
d
dr
O
x x x
d d
x x
dr = y d and ds = y d (9.27)
C=A x B
|B| sin
C = AB (9.28-a)
= |A||B| sin k (9.28-b)
i j k
99
= Ax Ay 0 = (Ax By Bx Ay ) k
(9.28-c)
Bx By 0
Area
|C| = |A||B| sin (9.28-d)
hence, the dierential area dxdy is then equal to the length of the vector resulting from the
cross product of drds and is equal to
# $
x x
d(area) = dxdy = det y y dd
(9.29)
J
32 The evaluation of the element stiness matrix involves dA. If we consider an innitesimal
element, of length dr and ds, at the vertex of an element, it has the boundaries of the element
as its sides. Then, from Eq. 9.28-d
dx.dy
detJ = sin (9.32)
d.d
Thus we observe that if is small or close to 180o , then det J will be very small, if the angle is
greater than 180 o , the determinant is negative (implying a negative stiness which will usually
trigger an error/stop in a FE analysis).
33 In general it is recommended that 30o < < 150o .
34 The inverse of the jacobian exists as long as the element is not much distorted or folds back
upon itself, Fig. ?? in those cases there is no unique relation between the coordinates.
35 It can be easily shown that for parallelograms, the Jacobian is constant, whereas for nonpar-
allelograms it is not.
36In general J is an indicator of the amount of element distorsion with respect to a 2x2 square
one. Some times it is constant, others it varies within the element.
9.2.2.1.1 Example: Jacobian Operators, (Bathe 1996) Determine the Jacobian op-
erators J for the following 2 dimensional elements. (Bathe 1982).
y
1
2 y 1
1
x 2
x
4 1 3/4
3 4
3 6 4
2
y
1
2
60
x 1
3 6 4
The coordinates are given by Eq. 9.4, the shape functions by Eq. 9.19, and the Jacobian
by Eq. 9.24-b.
Element 1:
1 1
x = (1 )(1 )x3 + (1 + )(1 )x4
4 4
1 1
+ (1 + )(1 + )x1 + (1 )(1 + )x2 (9.33-a)
4 4
1 1
= (1 )(1 )(3) + (1 + )(1 )(3)
4 4
1 1
+ (1 + )(1 + )(3) + (1 )(1 + )(3) (9.33-b)
4 4
1 1
y = (1 )(1 )y 3 + (1 + )(1 )y 4
4 4
1 1
+ (1 + )(1 + )y 1 + (1 )(1 + )y 2 (9.33-c)
4 4
1 1
= (1 )(1 )(2) + (1 + )(1 )(2)
4 4
1 1
+ (1 + )(1 + )(2) + (1 )(1 + )(2) (9.33-d)
4 4
3 0
[J] = (9.33-e)
0 2
Element 2:
1 1
x = (1 )(1 )((3 + 1/(2 3))) + (1 + )(1 )(3 1/2 3))
4 4
1 1
+ (1 + )(1 + )(3 + 1/2 3)) + (1 )(1 + )((3 1/2 3))) (9.34-a)
4 4
1 1
y = (1 )(1 )(2) + (1 + )(1 )(2)
4 4
1 1
911
Element 3:
1 1
x = (1 )(1 )(1) + (1 + )(1)
4 4
1 1
+ (1 + )(1 + )(1) + (1 )(1) (9.35-a)
4 4
1 1
y = (1 )(1 )(3/4) + (1 + )(1 )(3/4)
4 4
1 1
+ (1 + )(1 + )(5/4) + (1 )(1 + )(1/4) (9.35-b)
4 4
1 4 (1 + )
[J] = (9.35-c)
4 0 (3 + )
37 For a quadratic quadrilateral element, there are two possibilities, Fig. 9.9.
4 4
7 7
3 3
8 8 9
6 6
1 1
5 5
2 2
38 The Pascal triangle, Fig. 9.10, will be later used to justify the choice of terms in the
displacement eld of isoparameteric elements.
u = a1 + a2 x + a3 y + a4 x2 + a5 xy + a6 y 2 + a8 x2 y + a9 xy 2 (9.36)
0 1 2 3
40In this formulation, the x2 y 2 term is missing and the 8 terms in the assumed polynomial
expansion correspond the 8 nodes (4 corner and 4 midside).
Constant term 1
Quintic terms 3 2 2 3
Sixtic terms 3 3
y,v
913
4 7 4 7 3
3
1
6 8 6
8
1
5 2
1 1 5 2
x,u 1 1
Figure 9.11: Serendipity Isoparametric Quadratic Finite Element: Global and Parent Element
% + 2i&) (1 + i ) (i + i 1) i = 1, 2, 3, 4
1
Ni = 4 (1
Ni = 1
2 1 (1 % + i&) i = 5, 7 (9.37)
1
Ni = 2 (1 + i ) 1 + 2 i = 6, 8
i Ni Ni, Ni,
42 The shape functions for the corner and midisde nodes are shown in Fig. 9.12.
43 If we were to follow a similar procedure to the one adopted to extract the bilinear shape
fuctions, we would obtain 9 shape functions, which must in turn correspond to 9 (rather than
1 1
0.8
0.5
0.6
0
0.4
0.5
0.2
1 0
1 1
0.5 1 0.5 1
0 0.5 0 0.5
0 0
0.5 0.5
0.5 0.5
1 1 1 1
0 1 0.1
1
0.2
0.8
0.8
0.6
0.6 0.6
0.4
0.5
0.4 0.4
0.9
0.2 0.2
0.4 0.4
8) nodes.
44 In this element, the displacement eld is given by
u = a1 + a2 x + a3 y + a4 x2 + a5 xy + a6 y 2 + a8 x2 y + a9 xy 2 + a12 x2 y 2 (9.38)
0 1 2 3 4
Serendipity
45 All the quadratic terms are present, hence there are 9 terms in the polynomial expansion,
functions, yielding
Ni % + i&) (1 + i ) (i + i 1)
= 14 (1 i = 1, 2, 3, 4
Ni = 12 1 2 (1% + 2i&) i = 5, 7
(9.39)
Ni = 2 (1 + i ) 1
1
i = 6, 8
N9 = (1 2 )(1 2 ) i=9
49 Q9 elements perform much better than the Q8 if edges are not parallel or slightly curved.
915
50 The shape functions for the corner and midisde nodes are shown in Fig. 9.13.
0.5 0.8
0.6
0
0.4
0.5 0.2
0
1
1 0.2
1
0.5 1 0.5 1
0 0.5 0.5
0
0 0
0.5 0.5
0.5 0.5
1 1 1 1
1 1
0.8 0.8
0.6 0.6
0.6
0.4 0.4
0.6
0.2 0.2
0.2
0 0
1 0.1
0.8 0.4
1
0.6 0.6
0.8 0.2
0.4
0.8
0.6 0.2
0.4 0
0.9
0.2 0.2
0.4
0 0.7
1
0.6 0.5
0.5 1
0 0.5
0.8 0.3
0
0.5
0.5
1
1 1 1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
4 7 3
8 9 6
1 5 2
55For the six noded triangle, the partial derivatives of a variable with respect to x and y can
be expressed as, (Felippa 2000)
# $ )6 i Ni
L1 L2 L3 )i=1 L1
x x x x 6 Ni
= L i=1 i L2 (9.40)
y y
1 L
y
2 L
y
3
) 6 Ni
i=1 i L3
)6 )6 Ni )6 Ni L1 L1
1 1
Ni
)6 Ni )6 )6 i=1 L
i=1 L i=1 L2 x y x y
Ni
1 3
L2 L2 x x
x i xi N i
x i L3 = (9.42)
)i=1
6
L1
Ni )i=1
6
L2
Ni )i=1
6 Ni
x
L3
y
L3
x
y
y
y
i=1 yi L1 i=1 yi L2 i=1 yi L3 x y x y
y y
58 But x 1 1 x
x = y = 1, and x = y = y = x = 0 since x and y are independent coordinates.
Furthermore all entries in the rst row are equal to a constant (3 for the T6 element), and since
the corresponding right hand side, this row can be normalized, yielding the Jacobian matrix
for this element
L L1
1
1 1 1 0 0
) ) ) N L
x y
L
)i=1 xi L1i )i=1 xi L2i )i=1 xi L3i x2 y2 = 1 0
6 N 6 N 6
(9.43)
6 Ni 6 Ni 6 Ni L3 L3
y
i=1 i L1 y
i=1 i L2 y
i=1 i L3 x y
0 1
J
60 Next we invert the matrix and solve for the six triangular coordinates partials and substitute
in Eq. 9.40 which in turn will enable us to determine the B matrix in Eq. 9.23-a
Ni (L1 ,L2 ,L3 )
xx (L1 , L2 , L3 )
6 x 0
Ni (L1 ,L2 ,L3 ) ui
yy (L1 , L2 , L3 ) = 0 (9.45)
y vi
xy (L1 , L2 , L3 ) i=1 Ni (L1
y
,L2 ,L3 ) Ni (L1
x
,L2 ,L3 )
u
B=LN
61 Understanding numerical integration is not only essential for a proper integration of the
isoparameteric family of elements, but also helpful in understanding the Weighted Residual
methods (Chapter 11),
62A crucial aspect of isoparametric element formulation is the numerical integration which can
be expressed as
F()d or F(, )dd (9.46)
where the summations extend over all i and j, and Wi , Wij are weighting factors, and F(i )
and F(i , j ) are the matrices evaluated at the points specied in the arguments.
64 The matrices Rn are error matrices, which are in general not computed.
65 As shown above, F = BT .D.B for nite element stiness matrix evaluation, and each element
is integrated individually.
b
66 The integration of
a F ()d
b
is essentially based on passing a polynomial P () through given
values of F () and then use a P ()d as an approximation.
b b
F ()d P ()d (9.48)
a a
9.3.1 Newton-Cotes
68 In Newton-Cotes integration, it is assumed that the sampling points are equally spaced,
P( )
F( )
-1 0 1
b b
n n
b
P ()d = li ()F (i )d = li ()dF (i ) (9.50)
a a i=1 i=1 a
or
b )n (n)
Approximation P ()d = Wi F (i )
a
(n) b i=1
(n)
(9.51)
Weights Wi = a li ()d = (b a)Ci
(n)
where Ci are the weights of the Newton-Cotes quadrature for numerical integration
with n equally spaced sampling points.
n C0 C1 C2 C3 C4 Error
2 1
2
1
2 101 (b a)3 F II ()
3 1
6
4
6
1
6 103 (b a)5 F IV ()
4 1
8
3
8
3
8
1
8 103 (b a)5 F IV ()
5 7
90
32
90
12
90
32
90
7
90 106 (b a)7 F V I ()
69 Newton-Cotes constants, and corresponding reminder are shown in Table 9.3, (Bathe 1982).
70 It can be shown that this method permits exact integration of polynomial of order n 1,
and that if n is odd, then we can exactly integrate polynomials of order n. Hence we use in
general odd values of n,
711 For n = 2 over [1, 1], we select equally spaced points at 1 = 1 and 2 = 1 to evaluate
1 P ()d
2
P () = li ()F (i ) (9.52-a)
i=1
2 1
l1 () = = (1 ) (9.52-b)
1 2 2
1 1
l2 () = = (1 + ) (9.52-c)
2 1 2
1
(2) 1 1
W1 = l1 ()d = (1 )d = 1 (9.52-d)
1 2 1
1
(2) 1 1
W2 = l2 ()d = (1 + )d = 1 (9.52-e)
1 2 1
1 1
2
(2)
F ()d P ()d = Wi F (i ) = F (1) + F (1) (9.52-f)
1 1 i=1
3
P () = li ()F (i ) (9.53-a)
i=1
( 2 )( 3 ) 1
l1 () = = ( 1) (9.53-b)
(1 2 )(1 3 ) 2
( 1 )( 3 )
l2 () = = (1 + )( 1) (9.53-c)
(2 1 )(2 3 )
( 1 )( 2 ) 1
l3 () = = (1 + ) (9.53-d)
(3 1 )(3 2 ) 2
1 1
(3) 1 1
W1 = l1 ()d = ( 1)d = (9.53-e)
1 2 1 3
(3)
1
ISOPARAMETRIC ELEMENTS; 2nd GENERATION
1
4
W2 = l2 ()d = (1 + )( 1)d = (9.53-f)
1 1 3
1 1
(3) 1 1
W3 = l3 ()d = (1 + )d = (9.53-g)
1 2 1 3
1 1
3
(3) 1
F ()d P ()d = Wi F (i ) = [F (1) + 4F (0) + F (1)] (9.53-h)
1 1 3
i=1
73 In Gauss-Legendre quadrature, the points are not xed and equally spaced, but are selected
n
(n)
F ()d P ()d = Wi F (i ) (9.54)
1 1 i=1
(n)
however in this formulation both Wi and i are unknowns to be yet determined. Thus, we
have a total of 2n unknowns.
75At the integration points P (i ) = F (i ), however at intermediary points the dierence can
be expressed as
F () = P () + ()(0 + 1 + 2 2 + ) (9.55)
0 at =i ;i=1,2,,n
since we want the left side to be exactly equal to P () at the integration points, we dene
() = ( 1 )( 2 )...( n ) (9.56)
intermediary points.
77 Integrating
1 1
1
F ()d = P () + j () j d (9.57)
1 1 j=0 1
n1 1
1
j j
j () d = j () d + j () j d (9.58)
j=0 1 j=0 1 j=n 1
n1 1
F ()d P () + j () j d (9.59)
1 1 j=0 1
we observe that the rst integral on the right-hand side involves a polynomial of order n 1,
921
which would result in a set of n simultaneous equations of order n in terms of the unknowns
i , i = 0, 1, , n 1.
79 Back to Eq. 9.59
1 1
n 1
n
(n)
F ()d P () = F (i ) li ()d = Wi F (i ) (9.61)
1 1 i=1 1 i=1
or 1 )n (n)
Approximation 1 F ()d W F (i )
(n) 1i=1 i
Weights Wi = 1 li ()d (9.62)
1
Gauss Points j
1 () d = 0 j = 0, 1, , n 1
(n)
80 Integration points i and weight coecients Wi are given in Table 9.4, (Bathe 1982).
(n)
n i Wi Error
1 (2)
1 0 2 3F ()
2 1/
0 3 03
1/ 1 1 1
135 F
(4) ()
3 3/5 0 3/5 5/9 8/9 5/9 1
15,750 F (6) ()
Table 9.4: Integration Points and Weights for Gauss-Quadrature Formulaes Over the Interval
[1, 1]
81 The solutions (Gauss integration points) are equal to the roots of a Legendre polynomial
dened by
L0 () = 1
L1 () = (9.63)
k Lk1 () k Lk2 () 2kn
2k1 k1
Lk () =
and the n Gauss integration points are determined by solving Ln () = 0 for its roots i ,
i = 0, 1, , n 1.
82 The weighting functions are then given by
(n) 2(1 i2 )
Wi = (9.64)
[nLn1 (i )]2
9.3.2.2
ISOPARAMETRIC ELEMENTS; 2nd GENERATION
equations are
1
( 1 )( 2 ) 0 d = 0 (9.65-a)
1
1
( 1 )( 2 ) 1 d = 0 (9.65-b)
1
(9.65-c)
is accomplished by selecting m and n (not to be confused with the order of the polynomial)
integration points in the and directions
1 1 1 1
m
(m) (n)
F (.)dd P (, )dd = Wi Wj F (i , j ) (9.69)
1 1 1 1 i=1 j=1
and the total number of integration points will thus be m n, Fig. 9.16.
85 For the numerical integration over a triangle, the Gauss points are shown in Fig. 9.17, and
= + 0.6
= +1/ 3
= 1/ 3 = 0.6
= + 0.6
4 = +1/ 3 9
6
2 3
5 8
2
1 1 4
3 7
= 1/ 3 = 0.6
b
a
a b a
c
d
c
Table 9.5: Coordinates and Weights for Numerical Integration over a Triangle
9.4
ISOPARAMETRIC ELEMENTS; 2nd GENERATION
Figure 9.18: Extrapolation from 4-Node Quad Gauss Points (Felippa 1999)
hence any scalar quantity (such as xx ) whose values i is known at the Gauss element corners
can be interpolated through the usual bilinear shape functions now expressed in terms of and
1
( , ) = N1e N2e N3e N4e 2
(9.72)
3
4
Corner Gauss
925
Node Node
1 1 1 3 3 1 1/3 1/3 1 1
2 +1 1 +3 3 2 +1/3 1/3 +1 1
3 +1 +1 +3 +3 3 +1/3 +1/3 +1 +1
4 1 +1 3 + 3 4 1/ 3 +1/ 3 1 +1
where
1
N1e = (1 )(1 ) (9.73-a)
4
1
N2e = (1 + )(1 ) (9.73-b)
4
1
N3e = (1 + )(1 + ) (9.73-c)
4
1
N4e = (1 )(1 + ) (9.73-d)
4
Similarly, we can extrapolate to the cornersof the element. For corner 1, for instance, we
replace and in the preceding equations by 3. Doing that for the four corners, we obtain
1 1 + 12 3 12 1 12 3 12
1
2 2 1 + 2 3
1 1
2 1 2 3
1 1
2
= 1 1 3 (9.74)
3 12 1 + 12 3 12 3
2
4 12 1 12 3 12 1 + 12 3 4
As expected, the sum of each row is equal to one, and for stresses we replace by xx , yy ,
and xy
91 As we know, dierent nodal stresses will be obtained from adjacent elements. To obtain a
single value we can either take
1. Unweighted average of all the nodal stresses.
2. Weighted average of nodal stresses based on the relative sizes of the elements as determined
from their area through det(J).
92 In the nite element formulation, all loads must be replaced by an energy equivalent nodal
load.
93 We shall consider the following cases: nodal load, gravity, tractions, and thermal.
94 Gravity forces are equivalent to a body force/unit volume acting within the solid in the
Volume d 6
5
7
4
y
8
x 2
1 3
Direction of Gravity
97 Any element edge can have a distributed load per unit length in a normal and tangential
direction prescribed.
98 The variation of the distributed load is polynomial and its order can not exceed the order of
the element.
99 For the sake of consistency, loaded nodes are listed also counterclockwise.
100 First we determine the components of the distributed loads in the x and y directions by
considering the forces acting on an incremental length dS of the loaded edge, Fig. 9.20:
6
927
7
y P
4 t
dy
8
Pn
x 2 Pt
dx
1 3
Pn
The integration is again carried on numerically (along the edge) and the energy equivalent
nodal forces for node i are
ngaus
x y
Pxi = Ni pt pn Wj (9.84)
j=1
ngaus
x y
Pyi = Ni pn + pt Wj (9.85)
j=1
where the integration is carried on numerically along the edge. and note that x
and y
are
taken straight out of the Jacobian matrix.
103 Note that since integration is to be carried along the edge, we have used .
104 For adjacent elements, Fig. 9.21.
f0e = B D 0 d
T
BT 0 d (9.89)
e t
rized as follows.
But rst, it is assumed that this operation is to be performed in a function called stiff and
it takes as input arguments elcod, young, poiss, type, ndime, ndofn, ngaus. In turn it
will compute the stiness matrix KELEM of element ielem.
1. Retrieve element geometry and material properties for the current element
(a) Look up the sampling position of the current point (p , q ) (s, t) and their weights
(weigp)
(b) Call shape function routine sfr given p , q which will return the shape function Nie
N e N e
(sfr) and the derivatives i and i (deriv) at the point p , q
Nie Nie
(c) Call another subroutine (jacob), given Nie , and at point p , q will return
Nie Nie e
cartesian shape function derivatives x and y (cartd), the Jacobian matrix J
(jacm), its inverse Je1 (jaci), its determinant det Je (djac) and the x and y
coordinates all at the point p , q
Nie Nie
(d) Call strain matrix (bmatps) routine, given Nie , x , y , at p , q will return the
strain matrix Bei (bmat)
(e) Call a routine (dbmat) to evaluate De Bei (dbmat)
(f) Evaluate Bei T De Bej detJe integration weights and assemble them into the element
stiness matrix Keij
(g) Assemble De Be (smat into a stress array for later evaluation of stresses from the
nodal displacements.
idime, ndime
ISOPARAMETRIC ELEMENTS; 2nd GENERATION
%-----------------------------------------
% Extract element coordinates
%
% elcod = [ X1 X2 X3 . . . Xn]
% [ Y1 Y2 Y3 . . . Yn]
%-----------------------------------------
elcod = zeros(2,nnode);
for inode = 1:nnode
row = find(coord(:,1:1)==lnods(ielem,inode+3));
elcod(:,inode:inode) = coord(row:row,2:3);
end
%-----------------------------------------
% Constitutive matrix
%-----------------------------------------
D = dmat(young,poiss,type);
%-----------------------------------------
% Element stiffness matrix - element ielem
%-----------------------------------------
kelem = zeros(stifsize);
for igaus = 1:ngaus
s = posgp(igaus);
t = posgp(jgaus);
W = weigp(igaus)*weigp(jgaus);
[shape,deriv] = sfr(s,t,nnode);
[cartd,jacm,jaci,djac,xy] = jacob(shape,deriv,elcod);
[bmat,dbmat] = bmatps(shape,cartd,D);
kelem = kelem + bmat*dbmat*djac;
end
end
%----------------------------------------------------------------------
% Store element stiffness matricies in as a stack in a single matrix :
% kelem(1)
% KELEM = :
% kelem(nelem)
%----------------------------------------------------------------------
startrow = stifsize*ielem - nrowcount;
endrow = ielem*stifsize;
KELEM(startrow:endrow,:) = kelem;
end
t = toc;
fprintf(1,Elapsed time for this operation =%3.4fsec\n\n,t)
9.6.2.2 dmat.m
function D = dmat(young,poiss,type)
%-----------------------------------------------------------------------------------
% This function calculates the constitutive matrix for an element
%-----------------------------------------------------------------------------------
% VARIABLES
%-----------------------------------------------------------------------------------
% young Youngs modulus/Modulus of elasticity
% poiss Poissons ratio
% E Modulus of elasticity
% type type of problem - plain stress = 1, plain strain = 2
% D Constitutive matrix returned by function
%-----------------------------------------------------------------------------------
tic
%fprintf(Calculating element constitutive matrix\n)
E = young;
v = poiss;
%--------------------------
% Plain stress
%--------------------------
if type == 1.0
D = (E/(1-v^2))*[ 1 v 0;
v 1 0;
0 0 (1-v)/2];
%--------------------------
% Plain strain
%--------------------------
else
D = E/((1+v)*(1-2*v))*[ 1-v v 0 ;
933
v 1-v 0 ;
0 0 (1-2*v)/2];
end
t = toc;
%fprintf(1,Elapsed time for this operation =%3.4fsec\n\n,t)
9.6.2.3 sfr.m
function [shape,deriv] = sfr(s,t,nnode)
%-----------------------------------------------------------------------------------
% This function calculates the shape function and derivative for the current node
%-----------------------------------------------------------------------------------
% VARIABLES
%-----------------------------------------------------------------------------------
% shape Shape function returned by function
% deriv Derivative of shape function returned by function
% nnode Number of nodes per element
% s Natural coordinate (xi) of sampling point - horizontal
% t Natural coordinate (eta) of sampling point - vertical
%-----------------------------------------------------------------------------------
tic
%fprintf(Calculating shape functions and derivatives\n)
%-------------------
% Q9 Element
%-------------------
if nnode == 9
N9 = (1-s^2)*(1-t^2);
N8 = .5*(1-s)*(1-t^2) -.5*N9;
N7 = .5*(1-s^2)*(1+t) -.5*N9;
N6 = .5*(1+s)*(1-t^2) -.5*N9;
N5 = .5*(1-s^2)*(1-t) -.5*N9;
N4 = .25*(1-s)*(1+t) -.5*N7 - .5*N8 - .25*N9;
N3 = .25*(1+s)*(1+t) -.5*N6 - .5*N7 - .25*N9;
N2 = .25*(1+s)*(1-t) -.5*N5 - .5*N6 - .25*N9;
N1 = .25*(1-s)*(1-t) -.5*N5 - .5*N8 - .25*N9;
dN9ds = -2*s*(1-t^2);
dN9dt = -2*t*(1-s^2);
dN8ds = -.5*(1-t^2) -.5*dN9ds;
dN8dt = -t*(1-s) -.5*dN9dt;
dN7ds = -s*(1+t) -.5*dN9ds;
dN7dt = .5*(1-s^2) -.5*dN9dt;
dN6ds = .5*(1-t^2) -.5*dN9ds;
dN6dt = -t*(1+s) -.5*dN9dt;
dN5ds = -s*(1-t) -.5*dN9ds;
dN5dt = -.5*(1-s^2) -.5*dN9dt;
dN4ds = -.25*(1+t) -.5*dN7ds - .5*dN8ds - .25*dN9ds;
dN4dt = .25*(1-s) -.5*dN7dt - .5*dN8dt - .25*dN9dt;
dN3ds = .25*(1+t) -.5*dN6ds - .5*dN7ds - .25*dN9ds;
dN3dt = .25*(1+s) -.5*dN6dt - .5*dN7dt - .25*dN9dt;
dN2ds = .25*(1-t) -.5*dN5ds - .5*dN6ds - .25*dN9ds;
dN2dt = -.25*(1+s) -.5*dN5dt - .5*dN6dt - .25*dN9dt;
dN1ds = -.25*(1-t) -.5*dN5ds - .5*dN8ds - .25*dN9ds;
dN1dt = -.25*(1-s) -.5*dN5dt - .5*dN8dt - .25*dN9dt;
deriv = [dN1ds dN2ds dN3ds dN4ds dN5ds dN6ds dN7ds dN8ds dN9ds;
dN1dt dN2dt dN3dt dN4dt dN5dt dN6dt dN7dt dN8dt dN9dt];
%-------------------
% Q8 Element
%-------------------
elseif nnode == 8
N8 = .5*(1-s)*(1-t^2);
N7 = .5*(1-s^2)*(1+t);
N6 = .5*(1+s)*(1-t^2);
N5 = .5*(1-s^2)*(1-t);
N4 = .25*(1-s)*(1+t) -.5*N7 - .5*N8;
N3 = .25*(1+s)*(1+t) -.5*N6 - .5*N7;
N2 = .25*(1+s)*(1-t) -.5*N5 - .5*N6;
N1 = .25*(1-s)*(1-t) -.5*N5 - .5*N8;
dN8ds = -.5*(1-t^2);
dN8dt = -t*(1-s);
dN7ds = -s*(1+t);
dN7dt = .5*(1-s^2);
dN6ds = .5*(1-t^2);
dN6dt = -t*(1+s);
dN5ds = -s*(1-t);
dN5dt = -.5*(1-s^2);
dN4ds = -.25*(1+t) -.5*dN7ds - .5*dN8ds;
dN4dt = .25*(1-s) -.5*dN7dt - .5*dN8dt;
dN3ds = .25*(1+t) -.5*dN6ds - .5*dN7ds;
dN3dt = .25*(1+s) -.5*dN6dt - .5*dN7dt;
dN2ds = .25*(1-t) -.5*dN5ds - .5*dN6ds;
dN2dt = -.25*(1+s) -.5*dN5dt - .5*dN6dt;
dN1ds = -.25*(1-t) -.5*dN5ds - .5*dN8ds;
dN1dt = -.25*(1-s) -.5*dN5dt - .5*dN8dt;
%-------------------
% Q4 Element
%-------------------
else
N4 = .25*(1-s)*(1+t);
N3 = .25*(1+s)*(1+t);
N2 = .25*(1+s)*(1-t);
N1 = .25*(1-s)*(1-t);
dN4ds = -.25*(1+t);
dN4dt = .25*(1-s);
dN3ds = .25*(1+t);
dN3dt = .25*(1+s);
dN2ds = .25*(1-t);
dN2dt = -.25*(1+s);
dN1ds = -.25*(1-t);
dN1dt = -.25*(1-s);
end
t = toc;
%fprintf(1,Elapsed time for this operation =%3.4fsec\n\n,t)
9.6.2.4 jacob.m
function [cartd,jacm,jaci,djac,xy] = jacob(shape,deriv,elcod)
%-----------------------------------------------------------------------------------
% This function calculates the cartesian shape function derivatives
% the Jacobian matrix, Jacobian inverse and Jacobian determinant
%-----------------------------------------------------------------------------------
% VARIABLES
%-----------------------------------------------------------------------------------
% shape Shape function at current point
% deriv Derivative of shape function at current point
% cartd Cartesian shape function derivatives returned by function
% jacm Jacobian matrix returned by function
% jaci Jacobian matrix inverse returned by function
% djac Determinant of Jacobian matrix returned by function
% xy x and y coordinates at the current point in the element
%-----------------------------------------------------------------------------------
tic
%fprintf(Calculating Jacobian matrix\n)
%----------------------------------------------------------------
% The cartesian shape function derivatives {cartd} are given by:
%
% {dN/dx} -1 {dN/ds}
% {cartd} = { } = [J] { }
% {dN/dy} {dN/dt}
% Start by calculating Jacobian [J] = jacm:
% T T
% [dX/ds dY/ds] [{dN/ds}*{x} {dN/ds}*{y}]
% [J] = [ ] = [ T T ]
% [dX/dt dY/dt] [{dN/dt}*{x} {dN/dt}*{y}]
%----------------------------------------------------------------
jacm = deriv*elcod;
jaci = inv(jacm);
djac = det(jacm);
cartd = jaci*deriv;
xy = elcod*shape;
t = toc;
%fprintf(1,Elapsed time for this operation =%3.4fsec\n\n,t)
9.6.2.5 bmatps.m
function [bmat,dbmat] = bmatps(shape,cartd,D)
%-----------------------------------------------------------------------------------
% This function calculates the strain matrix B
%-----------------------------------------------------------------------------------
% VARIABLES
%-----------------------------------------------------------------------------------
% shape Shape function at current point
% cartd Cartesian shape function derivatives
% bmat Strain matrix returned by function
% dbmat Strain matrix * constitutive matrix D
%-----------------------------------------------------------------------------------
tic
%fprintf(Calculating strain matrix [B]\n)
numcols = 2*length(cartd);
bmat = zeros(3,numcols);
cartdcol = 0;
for ibmatcol = 1:2:numcols
cartdcol = cartdcol+1;
bmat(:,ibmatcol:ibmatcol+1) = [cartd(1,cartdcol) 0 ;
0 cartd(2,cartdcol);
cartd(2,cartdcol) cartd(1,cartdcol)];
end
dbmat = D*bmat;
t = toc;
%fprintf(1,Elapsed time for this operation =%3.4fsec\n\n,t)
meshc(X,Y,N1)
print -deps2 shap8-1.eps
c=contour(X,Y,N1);
clabel(c)
print -deps2 shap8-1-c.eps
meshc(X,Y,N8)
print -deps2 shap8-8.eps
c=contour(X,Y,N8);
clabel(c)
print -deps2 shap8-8-c.eps
N1=N1-0.25*N9;
meshc(X,Y,N1)
print -deps2 shap9-1.eps
c=contour(X,Y,N1);
clabel(c)
print -deps2 shap9-1-c.eps
N8=N8-0.5*N9;
meshc(X,Y,N8)
print -deps2 shap9-8.eps
c=contour(X,Y,N8);
clabel(c)
937
meshc(X,Y,N9)
print -deps2 shap9-9.eps
c=contour(X,Y,N9);
clabel(c)
print -deps2 shap9-9-c.eps
Chapter 10
where N denotes the number of nodes, dim is the problem dimension and R is the number
of rigid body modes. In Section 10.3, it will be shown that this formulation guarantees the
satisfaction of the BB condition.
The polynomial orders of the eld approximations are given in Table 10.1.
In general case a variable x is interpolated over a nite element using the expression:
Nen
x= i xi (10.2)
i
Draft
102
ELEMENT FORMULATION and STRAIN RECOVERY in HW
FORMULATION
where Nen is the number of element nodes, i is an interpolation function associated with node
i, and xi is the value of variable x at element node i. For the linear triangular element (T3)
the interpolation functions are:
i = li , i = 1, 2, 3 (10.3)
and for the six noded triangular element (T6) with three corner nodes and three mid-side nodes
the interpolation functions are:
where element nodes 1 to 3 indicate the element corner nodes and 4 to 6 are the mid-side nodes.
Symbols li denote the natural area coordinates of the element, which are related to the element
natural coordinates and by relations:
l1 = , l2 = , l3 = 1 (10.5)
For three-dimensional nite elements, the interpolation functions are similar. For the linear
tetrahedron T4 element they are:
i = li , i = 1, 2, 3, 4 (10.6)
and for the T10 element with four corner nodes and six mid-edge nodes they are dened
analogically to the six noded triangular T6 element as:
Similarly to the two-dimensional elements, symbols li denotes the volumetric natural coordi-
nates, which are again related to the element natural coordinates by relations:
l1 = , l2 = , l3 = , l4 = 1 (10.8)
10.2.1 C-lumping.
103
The inversion of C for the C-lumping (CL) technique is simplied by forming a diagonalized
C matrix. This lumped C matrix 1 is evaluated by the following expression:
CL = (I )dV (10.9)
V
where I is the identity matrix and is the shape function matrix. Since identical shape func-
tions are used for all three primary elds, the subscript at is no longer necessary. Replacing
C by C L Step 2 of Equation ?? reduces to:
n = C 1
L Eun (10.10)
At the element level, the lumped C matrix for the four node linear tetrahedron (T4) is:
5 0 0 0
0 5 0 0
C eL =
0
(10.11)
0 5 0
0 0 0 5
j+1
n = jn + C 1
L (Eun C n ).
j
(10.12)
where j = 0, 1, 2 . . . is the strain-iteration count. Note that this represents an internal iteration,
not to be confused with the MIM iteration of (??). The iteration process involves the nodal
strains in the whole mesh since Step 2 is equivalent to the least square t of the nodal based
strain eld to the strain eld derived from the displacement eld (Zienkiewicz and Taylor 1989).
For a four noded linear tetrahedral element (T4), the consistent matrix C is given at the
element level by:
2 1 1 1
1 2 1 1
C =
e
T dV = 1 1 2 1
(10.13)
Ve
1
Note: The diagonalized matrix C L and the consistent matrix
1 1 C 1described
2 in Section 10.2.2 and 10.2.3 are
determined in an analogous way as the standard lumped and consistent mass matrices in dynamics. The only
dierence is the exclusion of the weight density of the material which is replaced by unity.
jn = j+1
n jn = A() nj1 , (j 1) (10.14)
where:
A() = I C 1
L C (10.15)
where A() is a xed amplication matrix having a spectral radius = 45 . The spectral radius
is dened as the largest eigenvalue of amplication matrix A(). Since Equation 10.15 involves
a product of C and inverse of C L , the constants are cancelled out. By Banachs xed point
theorem (Haser and Sullivan 1991) it is necessary for the spectral radius to be less than 1
to ensure convergence of the iterative process given by Equation 10.12. Thus, this value of the
spectral radius indicates an error decay of 15 .
10.2.3 C-splitting.
A new iterative process was recently developed by (Cervenka, Keating and Felippa 1993) to solve
Step 2. This new technique guarantees faster convergence for linear triangular and tetrahedral
elements (T3 and T4). This technique is referred to as C-splitting (CS). This method splits
the consistent matrix C of Equation 10.13 into two matrices. One matrix is diagonalized and
the second is formed such that their algebraic sum is equivalent to the original C matrix:
C = CD + CR (10.16)
where:
C D = diag(C) CR = C CD (10.17)
is a splitting coecient controlling the splitting of the matrix C. Using this method Step
2 in Equation ?? is modied to:
j+1
n = C 1
D (Eun C R n ).
j
(10.18)
jn = j+1
n jn = A() nj1 , (j 1) (10.19)
A() = C 1
D C R, (10.20)
It is possible to select the coecient such that the spectral radius of the amplication matrix
is minimal. For four-node tetrahedron elements using = 32 , which as shown below minimizes
the spectral radius of the amplication matrix, C splits at the element level into the following
matrices:
3 0 0 0
0 3 0 0
C eD =
0 0 3 0
(10.21)
0 0 0 3
and
1 1 1 1
1 1 1 1
C eR =
1
(10.22)
1 1 1
1 1 1 1
Victor Saouma Finite Elements II; Solid Mechanics
Draft
10.3 Uniqueness and Existence of a Solution 105
Then the spectral radius is equal to 23 . Thus, CS has an error decay rate of 13 allowing
for a faster convergence than the SS method. For example, 10 steps of CS can be expected to
10
reduce the initial strain errors by ( 23 ) 0.0173 whereas 10 steps of SS would reduce those
10
errors by only ( 45 ) 0.1074.
This technique was investigated also for other low order element types. The best splitting
coecients and resulting spectral radii of the operator A() (Eq. 10.20) for other element
types are summarized in Table 10.2. The coecients for four node quadrilateral and eight
node brick element are however valid only for elements with parallel or almost parallel sides.
Therefore, the (SS) technique would be probably more reliable for these element types.
The third condition is satised as long as the material does not exhibit softening. This is always
guaranteed in the discrete crack approach, since softening is modeled only along the interface
elements which are not included in the mixed iterative solution.
If identical shape functions are used for all three elds, then the number of unknowns for
each eld is given by Equation 10.1, and the inequalities in the rst and second condition of
Equation 10.23 are always satised.
nu = N dim R n = N dim()
(10.25)
n = N dim() n= + nu = N dim() + N dim R
The rst condition is equivalent to the requirement that a nonzero vector of nodal displacements
must cause non-zero strain eld. It should be noted that the rigid body modes are excluded
from vector un . They would be the only modes allowed to produce a zero strain eld. In
this case, matrix B corresponds to that of a standard isoparametric triangular or tetrahedral
element, and will therefore satisfy this condition.
The second condition is also satised, since the shape functions in matrix are those of a
standard isoparametric element, and the uniqueness of the interpolation is guaranteed.
For higher order triangular and tetrahedral elements (i.e. 6 noded triangle and 10 noded
tetrahedron), it would seem preferable to select interpolation functions for strains and stresses,
which are one order lower than those for the displacements. This would correspond to the
mathematical relation between strains and displacements, since the strains are determined by
dierentiation of the displacement eld. For this formulation, there would be unknown dis-
placements, strains and stresses at each element corner node, but only displacement unknowns
at the midside element nodes. The strains and stresses would be interpolated using linear shape
functions and displacement using quadratic shape functions. Now, it is possible to show that
this formulation would not guarantee the satisfaction of the inequality in the rst condition of
Equation 10.23.
We consider a patch of two six-noded triangular element as shown in Figure 10.1. From
h h
5 5
2F
20
5
h h 5
F
8 and 9 noded isoparametric quadrilaterals; Pure Axial and Flexure
8 noded Isoparametric constant axial stress
the previous assumptions of linear stress and strain interpolation and quadratic displacement
interpolation, the unknown stresses and strains are only at the element corner nodes, while all
nodes have unknown displacements. The total number of stress and displacement unknowns is
nu = 9 2 3 = 15 > n = 4 3 = 12 (10.28)
and clearly the important inequality nu n is not satised and the existence and uniqueness
of a solution cannot be guaranteed. This should be contrasted by the previous formulation, in
which the same interpolation functions are used for all three elds, and the satisfaction of the
BB is guaranteed by Equation 10.25 and 10.27.
Chapter 11
WEIGHTED RESIDUAL
METHODS
11.1 Introduction
1 Thus far, the nite element method was derived from a variational principle through the
Rayleigh-Ritz method.
2 In Elasto-Statics, the variational principle used is the minimization of the total potential
energy, and the resulting weak formulation is equivalent to the equation of equilibrium (strong
form).
3 In many branches of physical sciences, a variational principle may not be available, and all
Symmetric (L[u])vd = (L[v])ud
(11.2)
Positive Denite (L[u])ud > 0
where u and v are any functions which satisfy essential and natural boundary conditions.
9 As an example, let us consider the steady state response of an axial bar. The governing
Hence 2
L[] = EA x
2; = u; g = 0;
(11.4)
B1 = 1; q1 = 0; B2 = EA x ; q2 = P ;
10To determine whether the operator L is symmetric and positive denite, we consider P = 0,
and integrating by part twice1 we obtain
l
l
2u u l u v
EA 2 vdx = EA v + EA dx (11.5-a)
0 x x 0 0 x x
l
u l v l 2v
= EA v + EAu EA 2 udx (11.5-b)
x x x 0 0 0
where w is an arbitrary weight function2 . We note that through this equation, the governing
113
where ai are the unknown parameters, and (x) are prespecied trial functions3 of x.
13 Substituting the previous equation into Eq. 11.8
b
w(Lapp + g)dx = 0 (11.10)
a
Since app will in general not satisfy the dierential equation exactly, then
Lapp + g = R (11.11)
b
wR = 0 (11.12)
a
where R(x) is a measure of the error, and is called the Residual; We thus impose the require-
ment that weights and residuals be orthogonal functions4 .
14 The residual clearly depends on the coecients ai , thus our objective is to select the weight
w(x) in such a way that the weighted residuals w(x)R(x) over a b is zero.
15 Arbitrary weight functions are selected rst
and Wi (x) are predened functions of x, ci are parameters. Hence, we can rewrite Eq. 11.8 as
b
T
c WT Rdx = 0 (11.14)
a
or
b b b b
W1 L(1 )dx W1 L(2 )dx W1 L(n )dx
W1 gdx
ab
a a
a
b b
b
W 2 L(1 )dx W 2 L(2 )dx W 2 L(n )dx W2 gdx
Ke = a a a ;f = a
.. .. .. ..
..
. . . .
.
b b
b
b
Wn L(1 )dx Wn L(2 )dx Wn L(n )dx
Wn gdx
a a a a
(11.18)
17Ke is thus an n by n matrix, and from the preceeding equation we can solve for thee unknown
coecients a.
18 The previous formulation was very general, and dierent techniques will vary according to
19In this method, the weight function w is based on Diracs delta function (x xi ) and
thus the residual is set to be exactly equal to zero at n points. This is achieved by adopting for
the weight function w the Diracs delta function (x xi ) where
if x = xi
(x xi ) = (11.19-a)
0 otherwise
(x xi )dx = 1 (11.19-b)
20 We subdivide the region of interest into n subregions, and within each one of them the
21 The integral of the square of the residuals are minimized with respect to ai .
I = R2 (x, y, z, a1 , , an )dx (11.20-a)
I
= 0; i = 1, 2, , n (11.20-b)
ai
W= (11.21)
i.e. the weight functions are equal to the trial functions. Since the trial functions are orthogonal
to the residuals (Eq. 11.8), then Eq. 11.15 becomes
b
T Rdx = 0 (11.22)
a
and
b b b b
1 L(1 )dx 1 L(2 )dx 1 L(n )dx
1 gdx
ab
a a
a
b b
b
2 L( 1 )dx 2 L( 2 )dx 2 L( n )dx 2 gdx
Ke = a a a ;f = a
.. .. .. ..
..
. . . .
.
b b
b
b
n L(1 )dx n L(2 )dx n L(n )dx
n gdx
a a a a
(11.23)
Formulate the nite element matrices by the Galerkin method for an an element of a string
subjected to a tension T (x), and to an external transverse force per unit length be p(x, t). The
mass per unit length is denoted by . Assume small displacements.
Solve for the fundamental frequency of vibration.
Solution:
1. We rst write Newtons equation of motion in the transverse direction (sigf y = ma) for
an element of the string
T (x) w(x, t) 2 w(x, t) w(x, t) 2 w(x, t)
T (x) + dx + 2
dx + p(x, t)dx T (x) = dx
x x x x t2
(11.24)
which reduces to
w(x, t) 2 w(x, t)
T (x) + p(x, t = (11.25)
x x t2
w
T(x) dx
T(x)+
p(x,t) x
w(x,t)
x 2
w(x,t) + w(x,t) dx
x x2
T(x)
w(x,t)
x x+dx x
2. In the case of free vibration p(x, t) = 0, and for constant tension T =cst, we obtain
where w = Nw
4. Integrate by parts
L L ! "L
NT,x T w,x dx NT wdx + NT T w,x 0 = 0 (11.28)
0 0
The last term is equal to zero because the shape functions at the corresponding nodes
are equal to zero. By analogy to the virtual displacement method, N corresponds to the
virtual displacement which must be equal to zero at the supports in order to satisfy the
essential boundary conditions.
5. Substituting w = Nw and w,x = N,x w we obtain
L L
T
N,x T N,x dx w + NT Ndx w = 0 (11.29)
0
0
k m
where
1
N= Lx
L
x
L and N,x = 1 1 (11.30)
L
6. Substituting
T 1 1
k = (11.31)
L 1 1
L 2 1
m = (11.32)
6 1 2
7. If we consider a two elements assemblage, and suppress the xed d.o.f (1 and 3), then
T L
2 w 2 + 4 w2 = 0 (11.33)
L 6
w 2 = A sin t (11.34)
w2 = 2 A sin t (11.35)
9. Substituting
T L
2 4 2 A sin t = 0 2 = 3 L
T
2 (11.36)
L 6
2 T
Note that the exact answer is w2 = 4L2
problems which does not have a variational principle, and thus the Galerkin method is the only
applicable technique. The second application will be the heat equation.
26 We will now apply the Galerkin method to the equation of elasticity and show that we will
retrieve the principle of virtual work which was derived from a variational principle. Hence, in
this particular case the two methods are indeed (as anticipated) equivalent.
27 Starting with the equilibrium equation (Eq. ??)
LT + b = 0 (11.37)
where = Lu and
xx
0 0
0
yy
x y z bx
zz
z ; =
LT = 0 0 0 ; b= b ; (11.38)
y x
xy
y
0 0
0
b z
z x y
xz
yz
xx xy
x + y + zxz + bx = 0
yx yy
x + y + zyz + by = 0 (11.39)
zx zy zz
x + y + z + bz = 0
29We multiply the rst equation in Eq. 11.39 by the arbitrary function wx and we integrate
over the volume5 :
xx xy xz
wx d + wx d + wx d + wx bx d = 0 (11.40)
x y z
tx = xx nx + xy ny + xz nz
ty = yx nx + yy ny + yz nz (11.42)
tz = zx nx + zy ny + zz nz
34 We next focus on the third term in the preceding equation. For w = wx wy wz T , then
119
wy wy wy
(Lw)T = wx
x ; y ;
wz
z ;
wx
y + x ;
wx
z + wz
x ; z + wz
y ; (11.46)
and
wy
(Lw)T = wx
x' xx + y + w z
z zz
(yy % & ' ( (11.47)
wy wy
+ w y
x
+ x xy + w z
x
+ wz
x xz + z + wz
y yz
36 But since the L operator is symmetric, from Eq. 11.2, the above equation can also be written
as
w (L )d =
T T T
w td + wT bd (11.49)
which is the weak form of the dierential equation of equilibrium subjected to the traction
(natural) boundary conditions.
37 The weight vector w is completely arbitrary. In the principle of virtual displacement, it
would correspond to the virtual displacement, compare this last equation with Eq. 5.11.
+ uT {G [D( 0 ) + 0 ] t}d = 0 (11.50)
11.4.2 FE Discretization
u = Nu (11.51)
where N are the shape functions, and u are the nodal known displacements. Note that this is
the application of Eq. 11.9-a
39 The Galerkin method uses the shape functions for the weight (functions Wi ) Hence, Eq.
11.13-a will be
w = Nc; wT = cT NT (11.52-a)
and
Lw = LNc = Bc (11.53)
40 Inserting the last two equations into the weak form, and recalling that the L operator is
(Note similarity with Eq. 11.14). Since c is an arbitrary matrix, we conclude that
BT d = NT td + NT bd (11.55)
This equation is applicable to both linear and nonlinear systems. The right hand side represent
the eect of the load.
41 Next we introduce a constitutive model, Eq. 3.90 which is rewritten in vector form
= D( 0 ) (11.56)
= Lu = LNu = Bu (11.57-a)
= DBu D0 (11.57-b)
44Thus, using the Galerkin approach, we have recovered the same stiness formulation as the
one previously obtained from a Variational Principle.
45 Whereas, there may not be any clearer advantage in using the Galerkin approach when a
variational principle is available, this may be the only formulation possible in the absence of
the second.
Chapter 12
FINITE ELEMENT
DISCRETIZATION OF THE
FIELD EQUATION
based formulation, we now turn our attention to a problem which (strictly speaking) could only
be handled by such an approach.
2 As a vehicle for such an application, we will consider the eld equation which was discussed
in Chapter 3.3.
4 Recalling (Eq. 3.98) that q = D, and that for steady state problems the right hand side
reduces to zero, and nally (for the mere sake of clarity) substituting by T , this equation
reduces to
div q Q = 0 (12.2)
L+g
Essential, Temperature T T = g
Natural, Flux q qn = qt n = qf (12.3)
Natural, Convection Flux c qc = h(T T )
6 Note that in a n, the convection heat loss can be considered as a negative heat source.
Draft
122 FINITE ELEMENT DISCRETIZATION OF THE FIELD EQUATION
(w)T qtdA = wqf td + wqn td wQtdA (12.8)
A q T A
where along q the ux (which may include convection) is known but the temperature is not, and
along T , the temperature is known but ux is not. This equation is valid for any constitutive
relation and is analogous to Eq. 11.49.
11 Applying Fourrier Law, we obtain
(w) DT tdA =
T
wqf td wqn td + wQtdA (12.9)
A q T A
12.2 FE Discretization
T = NT (12.11)
where N are the shape functions, and T are the nodal known temperatures.
RR H RR
1
A
div qdA =
qT nd A
(r)T qdA
T = T
x
T
y T (12.12)
16Note, this denition of the B matrix for scalar eld problems, should not be confused with
the one for vector problems, Eq. ??
N N
i i
0 0
x Ni 1 Ni
B= 0 y = J 0 (12.15)
Ni Ni Ni Ni
y x
12.2.1 No Convection
20 But since this expression should hold for any arbitrary cT matrix, we conclude that
Ke T = fbe + fQe
(12.21)
Ke = BT DBtdA (12.22)
A
fbe = N qf td
T
NT qn td (12.23)
q T
e
fQ = NT QtdA (12.24)
A
12.2.2
FINITE ELEMENT DISCRETIZATION OF THE FIELD EQUATION
Convection
= q + T + c (12.25)
The presence of the convection term aects only fb in Eq. 12.23 which now becomes
fb = NT qf td NT qn td NT qc d (12.26)
q T c
qc = h(T T ) (12.27)
or
qc = hNT hT (12.28)
It should be noted that convection occurs around the perimeter of an element, and conduction
occurs inside its volume.
24 Note the analogy between this load case and P- eects in structural engineering. As will
be seen in chapter 14, the geometric stiness matrix (KG ), which accounts for large deformation,
will be very similar to Kc .
o
T0 = 20 c
h K1 K2 K3
T
inf
1 2 3 4
1 2 3
o
20
q 1
0 11
00
0
1 00
11
c 0
1 00
11
0
1
0
1 00
11
00
11
0
1 00
11 Essential B.C.
Natural B.C. 0
1 00
11
The outer temperature is T0 = 20o C, convection heat transfer takes place on the inner surface
of the wall with the uid temperature T = 800o C, the lm coecient (or convection term)
h = 25 W/m2 .o C. We seek the temperature distribution across the wall.
Solution:
D=k (12.36)
Element Convection Matrix: Since convection occurs only at one point, 1, we just
add the convection term h = 25 to K1,1
1 2 3 4
66.67 + 25 66.67 0. 0.
66.67 266.67 200.00 0.
K =
0. 200.00 533.34 333.34
0. 0. 333.34 333.34
1 2 3 4
91.67 66.67 0. 0.
66.67 266.67 200.00 0.
= (12.38-a)
0. 200.00 533.34 333.34
0. 0. 333.34 333.34
0, and we do
not have a distributed convection (hence fb = NT qf d NT qn d +
q T
T
T N hd = 0) and we only have a point convection load
c
P1
(800)(25)
P2 0
= (12.39)
P 0
3
P4 ? P4 ?
A metallic n with thermal conductivity k = 360 W/m.o C, 0.1 cm thick and 10 cm long
extends from a plane whose temperature is 235 o C.
Tinf= 20o C
2
h = 9 W/m oc
0.1 cm
10 cm
1 2 Tinf= 20o C 3 4
We seek to determine the temperature distribution and amount of heat transfered from the n
to the air at 20o C where the lm coecient h=9 W/mo C.
25Using a three element discretization, and assuming the tip of the n to be isolated.
Solution:
1 2 3 4
129
1 1
360 1 1 + 1 1
K = (12.44-e)
0.033 1 1+1 1
1 1
1 2 3 4
1 1
360 1 2 1
= (12.44-f)
0.033 1 2 1
1 1
Convection Matrix The shape functions for the linear element are given by Eq. 12.35-a.
1 1+
N = 2
2 (12.45-a)
N1 N2
1
NT N = 2
1+ 1
2
1+
2
(12.45-b)
2
' (2
1 1+ 1
=
2 2 2
' (2 (12.45-c)
1+ 1 1+
2 2 2
1
T 1 2 1
N N = (12.45-d)
1 3 1 2
With respect to Fig. 12.1 heat transfer by convection occurs across the perimeter of the
t
B
ss
lo
at
he
n
tio
ec
nv
co
q x
dx
n (P = 2(B + t)); however, as noted above, we are considering a unit volume of the n
26 As for the conduction stiness matrix, the assembled convection stiness matrix will
be
1 2 3 4
2 1
2
(9)(3.33 10 ) 1 4 1
Kc = (12.49)
3(103 ) 1 4 1
1 2
1 2 3 4 1 2 3 4
1 1 2 1 T1
360 (9)(3.33 102 ) 1
1 2 1 + 4 1 T2 ?
=
0.033 1 2 1 3(103 ) 1 4 1 T3 ?
1 1 1 2 T4 ?
1
(20)(9)(3.33 102 ) 2
(12.53-a)
103
2
1
1 2 3 4 1 2 3 4
10, 800. 10, 800.
0. 0. 200. 100. 0. 0. T1 = 235
100.
10, 800. 21, 600. 10, 800. 0. + 400. 100. 0.
T2 ?
=
0. 10, 800. 21, 600. 10, 800. 0. 100. 400. 100. T3?
0. 0. 10, 800. 10, 800. 0. 0. 100. 200. T4 ?
P1 ?
12, 000.
(12.53-b)
12, 000.
6, 000.
11, 000. 10, 700. 0. 0.
235.
P 1 ?
10, 700. 22, 000. 10, 700. 0. T2 ? 12, 000.
0. 10, 700. 22, 000. 10, 700. T3 ? = 12, 000. (12.53-c)
0. 0. 10, 700 11, 000. T4 ? 6, 000.
Solving, we obtain
T2 208.77
o
T = 194.25 C (12.55)
3
T4 189.50
)
Heat loss in the n can be computed from H = e H e where H e = h(Tavg T )As sand the
surface area of each element is As = 2(1 0.0333)m2
12.4
FINITE ELEMENT DISCRETIZATION OF THE FIELD EQUATION
25 We now complement Table 3.2, by contrasting major governing equations in both scalar and
Scalar Vector
Conduction Solid Mechanics
I Dierential Equation
div q Q = 0 LT + b = 0 Balance/Equilibrium
T u State variable
rT " = Lu kinematic
q Flux vector
qn = qT n t = n Flux at boundary
q = DrT = D" Constitutive law
T on T u on u Essential BC
qn on q c t on t Natural BC
Z Z IIZ Weak Formulation
Z
T T
w (L )d = w td BT d = NT td Discretized Weak Form
Z Z
+ wT bd + NT bd
III Discretized Solution
T = NT u = Nu State variables
T = BT " = Bu
2 3 Flux
" # Ni
0
Ni x
x 6 Ni 7
B = Ni B = 4 0 y 5 State variable gradient
y Ni Ni
y x
(Ke + Kec )T = fbe + fQ
Z
e
+P Ke ue = fe + f0e + P Balance/Equilibrium
e T R T
K = B DBtdA Ke = B DBd Stiness matrix
Z A
e
NT qn td
Z T Z
e
fQ = NT QtdA fe = NT bd Body forces
A Ze
BT 0 d Initial stress
t
Chapter 13
TOPICS in STRUCTURAL
MECHANICS
13.1 Condensation/Substructuring
level, we may want to condense an internal node (such as in the Lagrangian element), at the
global level, this is closely associated with substructuring.
2 Let us consider kd = p and let d be partitioned into
dr
{d} = (13.1)
dc
where the subscripts r and c refer to retained and condensed d.o.f. respectively.
3 For substructuring, subscripts r and c refer to interface and interior d.o.f. Hence,
krr krc dr pr
= (13.2)
kcr kcc dc pc
or
k dr = p (13.3)
where
k = krr krc k1
cc kcr (13.4-a)
p = p r krc k1
cc pc (13.4-b)
4 Noting the common form of the above two equations, we deduce that condensation may
k = Tc kc (13.5-a)
p = Tc p (13.5-b)
Draft
132 TOPICS in STRUCTURAL MECHANICS
where
k1
cc kcr
c = (13.6)
I
5 k is a r by r matrix, whereae k was r+c by r+c (yet we still have to separately decompose
kcc which is c by c).
6 The patch test is a check which ascertains whether a patch of innitesimally small elements
subjected to constant strain reproduces exacly the constitutive behavior of the material through
correct stresses.
7 It has been argued that an element which passes the patch test satises the two essential
h h
5 5
2F
20
5
h h 5
F
8 and 9 noded isoparametric quadrilaterals; Pure Axial and Flexure
8 noded Isoparametric constant axial stress
9 Boundary nodes are then loaded by a consistently derived nodal loads appropriate to a state
of constant stress. This can be achieved by applying nodal load or nodal displacement. We
then verify that a state of constant stress is obtained inside the element. This test must be
repeated for each of the appropriate strains (xx , yy and xy ).
10 The patch test may pass by an individual element, but fail by an assemblage of such elements.
11 Failure of the patch test is also an indication of lack of stability. That is we have zero-energy
12 The stiness matrix is, by denition, singular due to the fact that equilibrium relations are
[d]1 [d]1 [B]T
[k] = (13.7)
[B][d]1 [B][d]1 [B]T
where B is the statics (or equilibrium) matrix, relating external nodal forces to internal forces;
d is a exibility matrix, and d1 is its inverse or reduced stiness matrix. We note that the
stiness matrix is obviously singular, since the second row is linearly dependent on the rst
one (through [B]) and thus, its determinent is equal to zero.
14 The reduced stiness matrix, which is the inverse of a exibility matrix, is not.
15 The internal strain energy stored in an element can be determined from
1
U = u {F}
2
1
= u[K] {u} (13.8)
2
16 If we consider a system where the load {F} applied to each node is proportional to the
or
([K] [I]) {u} = 0 (13.10)
degrees of freedom (or rows in [K]). To each eigenvalue i corresponds an eigenvector {u}i .
18 If the eigenvectors are normalized such that:
2. There should be as many zero eigenvalues as there are possible independent rigid body
motions (i.e. number of equations of equilibrium). For a two dimensional Lagrangian
element, there should be three zero eigenvalues, corresponding to two translations and
one rotation.
3. Too few zero eigenvalues is an indication of an element lacking the capability of rigid body
motion without strain.
6. Similar modes (such as exure in two orthogonal directions) will have identical eigenvalues
(for isotropic material).
7. When comparing the stiness matrices of two identical elements but based on dierent
formulations, the one with the lowest strain energy (tr[K] = i ) is best.
Rank: corresponds to the total number of linearly independent equations which is equal to the
order minus the number of rigid body motions.
Rank Deciency: would be equal to the total number of zero eigenvalues minus the rank.
21 For a square bilinear element the non-zero eigenvectors are shown in Fig. 13.2.
1 2 3 4 5
6 7 8
Figure 13.2: Eigenvectors Corresponding to a) Non-Zero and b) Zero Eigenvalues for a Square
Bilinear element
26 Reduced integration, i.e. using fewer Gauss points than required for an exact integration,
will cause the element to soften because higher order polynomial terms may vanish at the points
of lower order rule.
27 In an elastic body, the FE solution permits only those displacements which can be represented
by a linear combination of the shape functions. In general, the exact displacement eld will
dier from the assumed displacement, in other words the shape functions prevent the structure
from deforming the way it wants1 .
28 Given that the stiness formulation is too sti, (Ref. Sect. ??), the softening introduced by
a reduced numerical integration is benecial.
29 Recommended orders of integration for quadrilaterla elements are shown in Table 13.1.
Table 13.1: Full and Reduced Numerical Integrations for Quadrilateral Elements
.
30 If the element is articially softened, then there will be spurious zero energy modes, and
the stiness matrix will be rank decient, i.e. its rank is less than the number of element
d.o.f. minus the number of rigid body modes.
31 The independent displacement modes of a bilinear element are shown in Fig. 13.3. First
1
This is analogous to trucating a Fourrier series, which otherwise could have been an exact substitute to a
set of physical observations.
4 3
TOPICS in STRUCTURAL MECHANICS
1 2 3 4
1 2
5 6 7 8
three modes are rigid body motion, the next three are constant strains, and the last two are
bending modes.
32 Since the element does not provide any resistance to all spurious zero-energy modes (which
can nevertheless be used in linear combination with the other modes to describe the displace-
ment eld), then the eect of reduced integration is to soften the stiness of the FE model.
33 Should we adopt a reduced integration scheme, we will substantially reduce the computa-
tional eort, however we will forfeit some very special properties of the displacement method,
namely: a) Boundedness and b) monotonic convergence toward an exact solution.
34 An illustration of problems arising with reduced integration is illustrated for the following
elements
Q4 For a single point integration, the Gauss point is at the center. At the center of an element
under bending xx = yy = xy = 0, accordingly U = 0 for the corresponding modes 4
and 5 in Fig. 13.2, those spurious modes will disappear if n = 2.
Q8
35 In selective integration, we integrate the dierent strain terms with dierent orders of inte-
gration.
Figure 13.5: Rectangular Bilinear Element Subjected to Bending; Bilinear Element and Correct
Geometry
37 The shape functions for the bilinear element are given by Eq. 9.19
N1 (, ) = 1
4 (1 )(1 ); N2 (, ) = 1
4 (1 + )(1 );
(13.13)
N3 (, ) = 1
4 (1 + )(1 + ); N4 (, ) = 1
4 (1 )(1 + );
where = x/a and = y/b.
38 Imposing the displacement eld eld of Fig. 13.5 the bilinear element and correct displace-
Bilinear Exact
Element Formulation
u u u
v 0 (1 ) 2b + (1 2 ) 2a
2 au bu
xx ua ua
yy 0 ua
xy ub 0
39 From this table we note that parasitic shear contributes to the strain energy stored in the
element (which is equal to PT .d). Hence for a given imposed displacement, the forces would
be larger and M1 > M2 .
40 Computing the ratio of strain energies of the two elements, we obtain
M1 1 1 1 ' a (2
= + (13.14)
M2 1+ 1+ 2 b
hence, as a/b increases, the element locks (a term which will be explained in later chapters).
41 Hence, the unwanted shear strain that produces M1 > M2 is called parasitic shear.
Figure 13.6: Displacements Associated with Incompatible Modes for the Q6 Element
42 Comparing the displacement elds in Table 13.2 we observe that the bilinear element errs
from the exact solution by omitting the displacement mode associated with (1 2 ) and (1 2 ).
43 We can remedy to this situation by adding the missing modes as internal nodes
where the Ni are given by Eq. 13.13, and ai are nodeless d.o.f. The element now has 6 degrees
of freedom, and hence the element is referred to as Q6.
44The B matrix would have to be correspondingly adjusted and following its formulation, the
additional d.o.f. ai are eliminated through condensation (Sect. 13.1).
45 The Q6 element is incompatible (or nonconforming) due to the square distribution of u or
v along the edge which is dened by only two nodes, Fig. 13.6.
46 The softening in this case counters the inherent overstiness of the standard assumed dis-
placement formulation.
47 It can be shown that the Q6 element passes the patch test only if it is rectangular.
48To remedy to this deciency, we consider that portion of the B matrix associated with the
nodeless d.o.f. and label it Ba , and then the corresponding consistent nodal load vector is
e
pa =
e
BTa 0 d (13.16)
where 0 represent the initial stresses caused by the standard element formulation associated
with dr on the nodal d.o.f.
49 Because the standard isoparametric element passes the patch test, when 0 is constant,
without this additional load associated with the ai , then pea should also be zero for constant 0
or
1 1
BTa tdetJdd = 0 (13.17)
1 1
For general (non-rectangular elements) this equation is not satised and the patch test fails.
50 This can be articaily remedied in forming Ba and integrating by using J0 where the ja-
cobian is evaluated at = = 0 rather than at the Gauss quadrature point, and we use the
corresponding t0 for the thickness.
51 When such an operation is performed, the resulting element is called the QM6, and it works
almost as well as the quadratic element if rectangular, and is considerably more accurate than
the bilinear element.
52 Stresses are computed with all the d.o.f., including the ai . Better results are achieved if
element nodal loads pe that are associated with incompatible modes are set to zero during
recovery of the associated d.o.f. ai (following condensation).
53 Drilling d.o.f. are rotations at corner nodes. In the context of plane elements, they are
associated with parabolic displaced shapes of the edges, Fig. 13.7. They are of importance
in shells dened as an assemblage of at elements. Hence, the element can model exure and
membrane action.
54 Assuming rotation i and j at nodes i and j of an element side of length L, the end moments
of an immaginary beam will be
2EI
Mi = (2 i + j ) (13.18-a)
L
2EI
Mj = ( i + 2 j ) (13.18-b)
L
Hence, from M = Mi + (Mj Mi ) Lx and the vitual force method we can solve for the midside
edge-normal displacement
L
= ( j i ) (13.19)
8
55 Hence the edge displacement is caused by two parts: translation and parabolic distorsion:
u L s ui s uj Ls cos
= + + s(j i ) (13.20)
v L vi L vj 2L sin
13.5 Constraints
Chapter 14
GEOMETRIC NONLINEARITY
14.1 Introduction
Displacement
First Level elastic which excludes anly nonlinearities. This is usually acceptable for service
loads.
Elastic Critical load is usually determined from an eigenvalue analysis resulting in the
buckling load.
Secon-order elastic accounts for the eects of nite deformation and displacements, equi-
librium equations are written in terms of the geometry of the deformed shape, does not
account for material non-linearilties, may be able to detect bifurcation and or increased
stiness (when a member is subjected to a tensile axial load).
Draft
142 GEOMETRIC NONLINEARITY
First-order inelastic equilibrium equations written in terms of the geometry of the unde-
formed structure, accounts for material non-linearity.
2 This chapter will focus on elastic critical load determination. We will begin by reviewing the
derivation of some of the fundamental equations in stability analysis. We will examine both
the strong form, and the weak formulation.
imperfection is present.
5 At any location x along the column, the imperfection in the column compounded by the
Mz = P y (14.1)
d2 y P
2
y=0 (14.3)
dx EI
Letting k2 = P
143
7
EI , the solution to this second-order linear dierential equation is
1. y = 0 at x = 0, thus B = 0
2. y = 0 at x = L, thus
A sin kL = 0 (14.5)
9 This last equation can be satised if: 1) A = 0, that is there is no deection; 2) kL = 0, that
is no applied load; or 3)
kL = n (14.6)
P
% & 2
Thus buckling will occur if EI = n L or
n2 2 EI
P =
L2
10 The fundamental buckling mode, i.e. a single curvature deection, will occur for n = 1; Thus
2 EI
Pcr = (14.7)
L2
2 E
cr = % &2 (14.8)
L
r
where I = Ar 2 .
12 Note that buckling will take place with respect to the weakest of the two axis.
13In the preceding approach, the buckling loads were obtained for a column with specied
boundary conditons. A second order dierential equation, valid specically for the member
being analyzed was used.
14In the next approach, we derive a single fourth order equation which will be applicable to
any column regardelss of the boundary conditions.
15 Considering a beam-column subjected to axial and shear forces as well as a moment, Fig.
dv
14.3, taking the moment about i for the beam segment and assuming the angle dx between the
axis of the beam and the horizontal axis is small, leads to
dM (dx)2 dV dv
M M+ dx + w + V + dx P dx = 0 (14.9)
dx 2 dx dx
Victor Saouma Finite Elements II; Solid Mechanics
Draft
144
w
GEOMETRIC NONLINEARITY
w(x) w
M
P P V+ V dx
x P i
x
1111
0000 1111
0000 i i
0000
1111
0
1 1111
0000 v M+ M dx
P
0
1
0
1 x dx
P x j
P
0 y, u
1 dx j
Figure 14.3: Simply Supported Beam Column; Dierential Segment; Eect of Axial Force P
16 Neglecting the terms in dx2 which are small, and then dierentiating each term with respect
to x, we obtain
d2 M dV d2 v
P =0 (14.10)
dx2 dx dx2
17 However, considering equilibrium in the y direction gives
dV
= w (14.11)
dx
19 Substituting Eq. 14.11 and 14.12 into 14.10, and assuming a beam of uniform cross section,
we obtain
d4 v d2 v
EI 4 P 2 = w (14.13)
dx dx
P
20 Introdcing k2 = EI , the general solution of this fourth order dierential equation to any set
of boundary conditions is
v = C1 sin kx + C2 cos kx + C3 x + C4 (14.14)
21 If we consider again the stability of a hinged-hinged column, the boundary conditions are
v = 0, v,xx = 0 at x = 0
(14.15)
v = 0, v,xx = 0 at x = L
substitution of the two conditions at x = 0 leads to C2 = C4 = 0. From the remaining
conditions, we obtain
C1 sin kL + C3 L = 0 (14.16-a)
C1 k sin kl = 0
2
(14.16-b)
these relations are satised either if C1 = C3 = 0 or if sin kl = C3 = 0. The rst alternative
leads to the trivial solution of equilibrium at all loads, and the second to kL = n for n =
1, 2, 3 . For n = 1, the critical load is
2 EI
Pcr = (14.17)
L2
which was derived earlier using the lower order dierential equation.
145
22Next we consider a column with one end xed (at x = 0), and one end hinged (at x = L).
The boundary conditions are
v = 0, v,xx = 0 at x = 0
(14.18)
v = 0, v,x = 0 at x = L
But since cos kL can not possibly be equal to zero, the preceding equation can be reduced to
tan kL = kL (14.20)
which is a transcendental algebraic equation and can only be solved numerically. We are
essentially looking at the intersection of y = x and y = tan x, Fig. 14.4 and the smallest
10.0
8.0
6.0
4.0
2.0
0.0
-2.0
-4.0
-6.0
-8.0
-10.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0
Figure 14.4: Solution of the Tanscendental Equation for the Buckling Load of a Fixed-Hinged
Column
P
positive root is kL = 4.4934, since k2 = EI , the smallest critical load is
(4.4934)2 2
Pcr = 2
EI = EI (14.21)
L (0.699L)2
Note that if we were to solve for x such that v,xx = 0 (i.e. an inection point), then x = 0.699L.
23We observe that in using the higher order dierential equation, we can account for both
natural and essential boundary conditions.
24 Considering a uniform section prismatic element, Fig. ??, subjected to axial and exural
deformation (no shear), the Lagrangian nite strain-displacement relation is given by 3.26
1
xx = u,x + (u2,x + v,x
2 2
+ w,x ) (14.22)
2
thus, the total strain would be
2
du d v 1 dv 2
xx = y + (14.23)
dx
dx2 2 dx
Axial Flexure Large Deformation
25 We note that the rst and second terms are the familiar components of axial and exural
strains respectively, and the third one (which is nonlinear) is obtained from large-deection
strain-displacement.
26 The Strain energy of the element is given by
e 1
U = E2 d (14.24)
2 xx
28 Noting that
dA = A; ydA = 0; y 2 dA = I (14.26)
A A A
for y measured from the centroid, U e reduces to
# 2 2 4 2 $
2
1 du d v A dv du dv
Ue = E A +I 2
+ +A dx (14.27)
2 L dx dx 4 dx dx dx
A
% dv &4
We discard the highest order term 4 dx in order to transform the above equation into a
linear instability formulation.
29 Under the assumption of an independent prebuckling analysis for axial loading, the axial
load Px is
du
Px = EA (14.28)
dx
# 2 2 2 $
e 1 du d2 v dv
U = EA + EI + Px dx (14.29)
2 L dx dx2 dx
30We can thus decouple the strain energy into two components, one associated with axial and
the other with exural deformations
U e = Uae + Ufe (14.30-a)
2
e 1 du
Ua = EA dx (14.30-b)
2 L dx
# 2 2 2 $
1 d v dv
Ufe = EI 2
+ Px dx (14.30-c)
2 L dx dx
31 Recall, from Eq. 2.74 that for a functional in terms of two eld variables (u and v) with
higher order derivatives of the form
= F (x, y, u, v, u,x , u,y , v,x , v,y , , v,yy )dxdy (14.31)
there would be as many Euler equations as dependent eld variables, Eq. 2.75
F 2 F 2 2 F
u x u,x y u,y + x2 u,xx + xy u,xy + y 2 u,yy = 0
F F F
(14.32)
2 2 2 F
F
v x v,x
F F
y v,y
+ x F
2 v,xx +
F
xy v,xy + y 2 v,yy
= 0
d4 v d2 v
EI Px =0 (14.36)
dx4 dx2
joint displacements
v = Nu (14.37-a)
dv
= N,x u (14.37-b)
dx
d2 v
= N,xx u (14.37-c)
dx2
34 Substituting this last equation into Eq. 14.30-c, the element potential energy is given by
e = Ufe + W e (14.38-a)
1 1
= ue [ke ] {ue } + ue [kg ] {ue } u {P} (14.38-b)
2 2
where
[ke ] = EI {N,xx } N,xx dx (14.39)
L
and
[kg ] = P {N,x } N,x dx (14.40)
L
u1 v1 1 u2 v2 2
EA
L 0 0 EA
L 0 0
0 12EI 6EI
0 12EI 6EI
L3 L2 L3 L2
0 6EI 4EI
6EI 2EI
0 L
ke = EA L2 L L2 (14.41)
L 0 0 EA
0 0
L
0 12EI
L3
6EI
L2
0 12EI
L3
6EI
L2
0 6EI
L2
2EI
L 0 6EI
L2
4EI
L
which is the same element stiness matrix derived earlier in Eq. 8.7.
37 The geometric stiness matrix is given by
u1 v1 1 u2 v2 2
0 0 0 0 0 0
0 6 L
0 65 L
5 10 10
P
2
0 L 2 2
15 L 0 10 L30
L
kg = 10 (14.42)
L 0 0 0 0 0 0
L
0 65 10
L
0 6
10
5 Elements
Victor Saouma 2 Finite II; Solid Mechanics
0 L
10 L30 0 10 15 L
L 2 2
Draft
14.3 Elastic Instability; Bifurcation Analysis
ku = P (14.43)
where the element stiness matrix is expressed in terms of both the elastic and geometric
components)
k = ke + kg (14.44)
K = Ke + Kg (14.45)
we note that the structure becomes stier for tensile load P applied through Kg , and weaker
in compression.
40 We assume that conservative loading is applied, that is the direction of the load does not
follow the deected direction of the member upon which it acts.
41 In elastic instability, the intensity of the axial load system to cause buckling is yet unknown,
the incremental stiness matrix must rst be numerically evaluated using an arbitrary chosen
load intensity (since Kg is itself a function of P ).
42 For buckling to occur, the intensity of the axial load system must be times the initially
arbitrarily chosen intensity of the force. Note that for a structure, the initial distribution of P
must be obtained from a linear elastic analysis. Hence, the buckling load, P is given by
P = P (14.46)
43 Since the geometric stiness matrix is proportional to the internal forces at the start, it
follows that
Kg = Kg (14.47)
where Kg corresponds to the geometric stiness matrix for unit values of the applied loading
( = 1).
44 The elastic stiness matrix Ke remains a constant, hence we can write
(Ke + Kg )u P = 0
and for the displacements to tend toward innity (i.e buckling/bifurcation/instability), then
|Ke + Kg | = 0 (14.48)
|K1
g Ke + I| = 0 (14.49)
46 Alternatively, it can simply be argued that there is no unique solution (bifurcation condition)
to u.
47 The lowest value of , crit will give the buckling load for the structure and the buckling
loads will be given by
Pcrit = crit P (14.50)
48 The corresponding deformed shape is directly obtained from the corresponding eigenvector.
Solution:
1 2 3 4 5 6
EA
L 0 0 EA
L 0 0
0 12EI 6EI
0 12EI 6EI
L3 L2 L3 L2
0 6EI 4EI
6EI 2EI
0
k1e = EA L2 L L2 L
(14.51-a)
L 0 0 EA
0 0
L
0 12EI
L3 6EI
L2 0 12EI
L3 6EI
L2
0 6EI
L2
2EI
L 0 6EI
L2
4EI
L
4 5 6 7 8 9
1411
EA
L 0 0 EA
L 0 0
0 12EI 6EI
0 12EI 6EI
L3 L2 L3 L2
0 6EI 4EI
L2
6EI 2EI
0
k2e = EA L2 L L
(14.51-b)
L 0 0 EA
0 0
L
0 12EI
L3
6EI
L2
0 12EI
L3
6EI
L2
0 6EI
L2
2EI
L 0 6EI
L2
4EI
L
1 2 3 4 5 6
0 0 0 0 0 0
0 6 L
0 65 L
5 10 10
P
2
0 L 2 2
15 L 0 10
L
L30
k1g = 10 (14.52-a)
L 0 0 0 0 0 0
L
0 65 10
L
0 6
5 10
2
L
0 10 L30 0 10
L 2 2
15 L
4 5 6 7 8 9
0 65 L
10 0 65 L
10
2
L 2 2
0 10
L
L30
P
0 10 15 L
k2g = 0 0 0 0 0 0 (14.52-b)
L L
0 65 10
L
0 6
5 10
2
L
0 10 L30 0 10
L 2 2
15 L
The structures stiness matrices Ke and Kg can now be assembled from the element sti-
nesses. Eliminating rows and columns 2, 7, 8, 9 corresponding to zero displacements in the
column, we obtain
1 4 3 5 6
AL2 AL2
I I 0 0 0
AL2 2 AL2 0 0 0
EI
I I
Ke = 3 0 0 4L 2 6L 2L2 (14.53)
L
0 0 6L 24 0
0 0 2L2 0 8L2
and
1 4 3 5 6
0 0 0 0 0
0 0 0 0 0
P
0 2 2 L L2
Kg = 0 15 L (14.54)
L
0 0 L
10
12
30
0
10 5
L2 4 2
0 0 30 0 15 L
1 4 3 5 6
AL2 2
1 I AL 0 0 0
I
4 AL
2 2
I 2 AL
I 0 0 0
2 L4 1 L3 1 L4
3 0 0 4L 15
2 6L + 10 2
2L + 30 =0 (14.55)
EI EI EI
5 0 0 1 L3
6L + 10 EI 24 12 L2
5 EI 0
6 0 0 1 L4
2L2 + 30 EI 0 4 L4
8L2 15 EI
AL2 L2
introducing = I and = EI , the determinant becomes
1 4 3 5 6
1 0 0 0
4 2 0 0 0
% &
3 0 0 2 2 15 6L
+ 10& 2 + 30 = 0
(14.56)
%
5 0 0 6L + 10
12 2 5
% 0 &
6 0 0
2 + 30 0 4 2 15
(4.4934)2 (4.4934)2
Pcr = EI = EI = 5.0477 EI
L2
(14.58)
l2 (2L)2
and thus, the numerical value is about 2.6 percent higher than the exact one. The mathematica
code for this operation is:
(* Define elastic stiffness matrices *)
ke[e_,a_,l_,i_]:={
{e a/l , 0 , 0 , -e a/l , 0 , 0 },
{0 , 12 e i/l^3 , 6 e i/l^2 , 0 , -12 e i/l^3 , 6 e i/l^2 },
{0 , 6 e i/l^2 , 4 e i/l , 0 , -6 e i/l^2 , 2 e i/l },
{-e a/l , 0 , 0 , e a/l , 0 , 0 },
{ 0 , -12 e i/l^3 , -6 e i/l^2 , 0 , 12 e i/l^3 , -6 e i/l^2 },
{ 0 , 6e i/l^2 , 2 e i/l , 0 , -6 e i/l^2 , 4 e i/l }
}
ke1=N[ke[e,a,l,i]]
ke2=N[ke[e,a,l,i]]
{0 , 0 , 0 , 0 , 0 , 0 },
1413
Determine the buckling load for the following frame. Neglect axial deformation.
2 P P
1
0
0
1 3
11
00
u1
0
1
0
1 u1
I=200
I=50 6
10
I=100
1111
0000
0000
1111
15
11111111111111111111
00000000000000000000
Solution:
The element stiness matrices are given by
u1 2 0 0
20 1, 208
1, 208 96, 667
k1e = (14.59-a)
u1 2
GEOMETRIC NONLINEARITY
0 0
0.01 0.10
0.10 16.00
k1g = P
(14.59-b)
0 2 0 3
128, 890 64, 440
k2e = (14.59-c)
64, 440 128, 890
u1 3 0 0
47 1, 678
1, 678 80, 556
k3e = (14.59-d)
u1 3 0 0
0.01667 0.1
0.1 9.6
k3g = P
(14.59-e)
(Ke P Kg ) = 0 (14.60)
u1 2 3
(66.75) P (0.026666) (1, 208.33) P (0.1) (1, 678.24) P (0.1)
(1, 208.33) P (0.1) (225, 556.) P (16.) (64, 444.4) P (0) = 0 (14.61)
(1, 678.24) P (0.1) (64, 444.) P (0) (209, 444.) P (9.6)
The smallest buckling load amplication factor is thus equal to 2, 017 kips.
(* Initialize constants *)
a1=0
a2=0
a3=0
i1=100
i2=200
i3=50
l1=10 12
l2=15 12
l3=6 12
e1=29000
e2=e1
e3=e1
(* Define elastic stiffness matrices *)
ke[e_,a_,l_,i_]:={
{e a/l , 0 , 0 , -e a/l , 0 , 0 },
{0 , 12 e i/l^3 , 6 e i/l^2 , 0 , -12 e i/l^3 , 6 e i/l^2 },
{0 , 6 e i/l^2 , 4 e i/l , 0 , -6 e i/l^2 , 2 e i/l },
{-e a/l , 0 , 0 , e a/l , 0 , 0 },
{ 0 , -12 e i/l^3 , -6 e i/l^2 , 0 , 12 e i/l^3 , -6 e i/l^2 },
}
ke1=ke[e1,a1,l1,i1]
ke2=ke[e2,a2,l2,i2]
ke3=ke[e3,a3,l3,i3]
(* Define geometric stiffness matrices *)
kg[l_,p_]:=p/l{
{0 , 0 , 0 , 0 , 0 , 0 },
{0 , 6/5 , l/10 , 0 , - 6/5 , l/10 },
{0 , l/10 , 2 l^2/15 , 0 , - l/10 , - l^2/30 },
{0 , 0 , 0 , 0 , 0 , 0 },
{0 , -6/5 , - l/10 , 0 , 6/5 , - l/10 },
{0 , l/10 , - l^2/30 , 0 , - l/10 , 2 l^2/15 }
}
kg1=kg[l1,1]
kg3=kg[l3,1]
(* Assemble structure elastic and geometric stiffness matrices *)
ke={
{ ke1[[2,2]]+ke3[[2,2]] , ke1[[2,3]] , ke3[[2,3]] },
{ ke1[[3,2]] , ke1[[3,3]]+ke2[[3,3]] , ke2[[3,6]] },
{ ke3[[3,2]] , ke2[[6,3]] , ke2[[6,6]]+ke3[[3,3]] }
}
kg={
{ kg1[[2,2]]+kg3[[2,2]] , kg1[[2,3]] , kg3[[2,3]] },
{ kg1[[3,2]] , kg1[[3,3]] , 0 },
{ kg3[[3,2]] , 0 , kg3[[3,3]] }
}
(* Determine critical loads in terms of p (note p=1) *)
p=1
keigen=Inverse[kg] . ke
pcrit=N[Eigenvalues[keigen]]
modshap=N[Eigensystems[keigen]]
49 From Eq. 14.44 it is evident that since kg depends on the magnitude of Px , which itself may
[Ke + Kg ] u = P (14.62)
Determine the midspan displacement and member end forces for the beam-column shown
below in terms of Px ; The concentrated force is 50 kN applied at midspan, E=2 109 kN/m2
and I=2 103 m4 .
50 kN
1
0
0
1
0
1
0
1
80,000 kN1
0 0
1 0
1 80,000 kN
1
0 0
1
0
1 0
1
0000000000000000
1 1111111111111111
111111111111111000000000000000
6 m 6 m 0
Solution:
Using two elements for the beam column, the only degrees of freedom are the deection and
rotation at midspan (we neglect the axial deformation).
The element stiness and geometric matrices are given by
0 0 0 0 v1 2
0 0 0 0 0 0
0 222, 222. 666, 666. 0 222, 222. 666, 666.
0 666, 666. 2, 666, 666 0. 666, 666. 1, 333, 333
[Ke ] =
1
0
(14.65)
0 0 0 0 0
0 222, 222. 666, 666. 0 222, 222. 666, 666.
0 666, 666. 1, 333, 333 0. 666, 666. 2, 666, 666
0 v1 2 0 0 0
0 0 0 0 0 0
0 222, 222. 666, 666. 0 222, 222. 666, 666.
0 666, 666. 2, 666, 666 0. 666, 666. 1, 333, 333
[Ke ] =
2
0
(14.66)
0 0 0 0 0
0 222, 222. 666, 666. 0 222, 222. 666, 666.
0 666, 666. 1, 333, 333 0. 666, 666. 2, 666, 666
0 0 0 0 v1 2
0 0 0 0 0 0
0 16, 000 8, 000 0 16, 000 8, 000
0 8, 000 64, 000 0 8, 000 16, 000
[Kg ] =
1
0
(14.67)
0 0 0 0 0
0 16, 000 8, 000 0 16, 000 8, 000
0 8, 000 16, 000 0. 8, 000 64, 000
0 v1 2 0 0 0
0 0 0 0 0 0
0 16, 000 8, 000 0 16, 000 8, 000
0 8, 000 64, 000 0 8, 000 16, 000
[Kg ] =
2
0
(14.68)
0 0 0 0 0
0 16, 000 8, 000 0 16, 000 8, 000
0 8, 000 16, 000 0. 8, 000 64, 000
Victor Saouma Finite Elements II; Solid Mechanics
Draft
14.4 Second-Order Elastic Analysis; Geometric Non-Linearity
v1 2
412, 444. 0.
[K] = (14.69)
0. 5, 205, 330
Plf t
ulf t
Vlf t
vlf t
! 1 "
Mlf t 1 lf t
= [Ke ] + [Kg ]
Prgt
urgt
Vrgt
vrgt
Mrgt rgt
0 0 0 0 v1 2
0 0 0 0 0 ,0
0
0 206, 222. 0 206, 222. 658, 667.
658, 667. 0
0 658, 667.
260, 2670 0 658, 667. 1, 349, 330 0
= 0
0 0 0 0 0 0
0 206, 222
658, 667. 0 206, 222. 658, 667. 0.00012123
0 658, 667. 1, 349, 330 0 658, 667. 2, 602, 670 0
0
25.
79.8491
= (14.71-a)
0.
25.
79.8491
Note that had we not accounted for the axial forces, then
v1 0.0001125
= (14.72-a)
2 0
Plf t
0
Vlf t
25.
Mlf t 75.
= (14.72-b)
Prgt
0.
Vrgt
25.
Mrgt 75.
Plf t
0
GEOMETRIC NONLINEARITY
Vlf t
25.
Mlf t 70.8022
= (14.73-b)
Prgt
0.
V
25.
rgt
Mrgt 70.8022
We observe that the compressive force increased the displacements and the end moments,
whereas a tensile one stiens the structure by reducing them.
The Mathematica to solve this problem follows
(* Initialize constants *)
OpenWrite["mat.out"]
a1=0
a2=0
e=2 10^9
i=2 10^(-3)
i1=i
i2=i1
l=6
l1=l
l2=6
p=-80000 (* negative compression *)
load={-50,0}
(* Define elastic stiffness matrices *)
ke[e_,a_,l_,i_]:={
{e a/l , 0 , 0 , -e a/l , 0 , 0 },
{0 , 12 e i/l^3 , 6 e i/l^2 , 0 , -12 e i/l^3 , 6 e i/l^2 },
{0 , 6 e i/l^2 , 4 e i/l , 0 , -6 e i/l^2 , 2 e i/l },
{-e a/l , 0 , 0 , e a/l , 0 , 0 },
{ 0 , -12 e i/l^3 , -6 e i/l^2 , 0 , 12 e i/l^3 , -6 e i/l^2 },
{ 0 , 6e i/l^2 , 2 e i/l , 0 , -6 e i/l^2 , 4 e i/l }
}
ke1=N[ke[e,a1,l1,i1]]
ke2=N[ke[e,a2,l2,i2]]
(* Assemble structure elastic stiffness matrices *)
ke=N[{
{ ke1[[5,5]]+ke2[[2,2]], ke1[[5,6]]+ke2[[2,3]]},
{ ke1[[6,5]]+ke2[[3,2]], ke1[[6,6]]+ke2[[3,3]]}
}]
WriteString["mat.out",MatrixForm[ke1]]
WriteString["mat.out",MatrixForm[ke2]]
WriteString["mat.out",MatrixForm[ke]]
(* Define geometric stiffness matrices *)
kg[p_,l_]:=p/l {
{0 , 0 , 0 , 0 , 0 , 0 },
{0 , 6/5 , l/10 , 0 , - 6/5 , l/10 },
{0 , l/10 , 2 l^2/15 , 0 , - l/10 , - l^2/30 },
{0 , 0 , 0 , 0 , 0 , 0 },
{0 , -6/5 , - l/10 , 0 , 6/5 , - l/10 },
{0 , l/10 , - l^2/30 , 0 , - l/10 , 2 l^2/15 }
}
kg1=N[kg[p,l1]]
kg2=N[kg[p,l2]]
(* Assemble structure geometric stiffness matrices *)
kg=N[{
{ kg1[[5,5]]+kg2[[2,2]], kg1[[5,6]]+kg2[[2,3]]},
{ kg1[[6,5]]+kg2[[3,2]], kg1[[6,6]]+kg2[[3,3]]}
}]
1419
Analyse the stability of the following structure. Compare the axial force caused by the
coupled membrane/exural eects with the case where there is no interaction.
1,000
1
0
0
1
0000000000
1111111111 0
1
0000000000
1111111111
0000000000
1111111111 0
1 12
0000000000
1111111111
12
0000000000
1111111111 0
1
0000000000
1111111111
00000000001111111111
1111111111 0000000000
0
1 /8 /8 0
1
0
1 0
1
Solution:
In the following solution, we will rst determine the axial forces based on the elastic stiness
matrix only. Then, on the basis of those axial forces, we shall determine the geometric stiness
matrix, and solve for the displacements. Because of the non-linearity of the problem, we may
have to iterate in order to reach convergence. Following each analysis, we shall recompute the
geometric stiness matrix on the basis of the axial loads detemined from the previous iteration.
Note that convergence will be reached only for stable problems. If the method fails to
converge, it implies possible biurcation which could be caused by elastic displacements ap-
proaching L sin , due to either being too small, or E being too small (i.e not sti enough).
(*
1421
(*
GEOMETRIC NONLINEARITY
Solve
*)
Ks = Ke + Kg;
diseg = Inverse[Ks] . load;
normnew = Sqrt[ diseg . diseg ];
ratio = ( normnew-normold ) / normnew;
Print["Iteration ",N[iter],"; u1 ",N[u1],"; p1 ",N[p1]," ratio ",N[ratio]];
normold = normnew;
If[ Abs[ ratio ] < epsilon, Break[] ]
]
Print[" p1 ",N[p1]," up1 ",N[up1]," p1/up1 ",N[p1/up1]," ratio ",N[ratio]]
14.5 Summary
'
2
( % dv &2
2nd Order D.E. 4th Order D.E. x = du
dx y d v
dx2 + 1
2 dx
2 B.C. 4 B.C.
d2 y 1
2
dx2
P
EI=0 U= 2 E d
v = A sin kx B cos kx
d4 v d2 v
P
EI dx 4 =w
dx2 K= Ke Kg
v = C1 sin kx + C2 cos kx + C3 x + C4
P = (Ke + Kg )u
|Ke + Kg | = 0
Chapter 15
PLATES
Draft
1 This chapter will cover the transverse deformation of plates. The approach followed will be
Fundamental Relations
?
Dierential Equation
?
Variational Formulation
?
Finite Element Discretization
2 Considering an arbitrary plate, the stress are given by, Fig. 15.2, resultants per unit width
Draft
152 PLATES
are given by
t
2
Nxx = xx dz
2t
t
2
t
2
Membrane Force N = dz Nyy = yy dz
2t
2t
t
2
Nxy = xy dz
2t
t
2
Mxx = xx zdz
2t
t
(15.1-a)
2
t
2
Bending Moments M = zdz Myy = yy zdz
2t
2t
t
2
Mxy = xy zdz
2t
t
2
t
Vx = xz dz
2 2t
Transverse Shear Forces V = dz t
2t
2
Vy = yz dz
2t
3 Note that in plate theory, we ignore the eect of the membrane forces, those in turn will be
accounted for in shells.
4 The equation of equilibrium is derived by considering an innitesimal element tdx dy sub-
jected to an applied transverse load pz . We would have to consider three equations of equilib-
rium, Fig. 15.3:
Summation of Forces in the z direction
Vx Vy
dxdy + dxdy + pz dxdy = 0 (15.2)
x y
or
Vx Vy
+ + pz = 0 (15.3)
x y
x
- Vy
y ;
; 6 Mxy
;
; Myy
;
z ?V ; ; --
x
6 Mxx
; ;;
pz
-- ?
Mxy Mxx + @M@x dx ;
;
xx
Mxy + @M@x dx
xy
?
;
Myy + @M@y dy ;;
yy
Vx + @V@x dx
?Vy + @V@y dy
; y
x
Mxy + @M@y dy
xy
Mxy Myy
dxdy + dxdy Vy dxdy = 0 (15.4)
x y
or
Mxy Myy
+ Vy = 0 (15.5)
x y
Summation of Moments about the y axis
Myx Mxx
+ Vx = 0 (15.6)
y x
Note that the left matrix corresponds to LT where the 1 term has been substituted by 1
5 From Fig. 15.4 we have ve displacements u, v, w, xx and yy . However, two of the three
displacements, u, v can be expressed in terms of the xx and yy which are the rotations of the
plate middle surface. The third displacement being the transverse one w.
6 We will use a notation consistent with the traditional one adopted for plate bending, rather
than one consistent with the coordinate directions as used in nite element.
7 The middle surface will be assumed to remain unstrained, and that plane sections remain
plane. When shear deformations will later be taken into account, then the plane will not
necessarily remain normal to the middle surface.
8 Based on the plane section remaining plane assumption, we have,
u z 0 0 xx
v = 0 z 0 yy (15.8)
w 0 0 1 w
10 Since w corresponds to the transverse deection of the middle surface, and does not vary
with z, then zz = 0.
xx
zxx,x
zxx
yy zyy,y zyy
zz 0 0
= = (15.10)
xy
z( xx,y + yy,x )
2z xy
xz
(xx + w,x )
xz
yz (yy + w,y ) yz
Note that this equation assumes that the displacement w is small compared to the thickness of
the plate, and the rotation is small. Because the rotation is small, its square is negligible with
respect to unity, and hence the curvature
= x
(15.11)
(1 + 2 )3/2 x
But From Eq. 15.10 zz = 0, and we can neglect zz which is much smaller than the other
stresses. Hence, inverting the previous equation yields
xx 1 0
xx
1
yy E
0 0 yy
xy = 0 0 1 xy (15.14)
1 2
2
1
0
xz
xz
0 2
1
yz 0 yz
2
D
E
Note that the shear modulus is = 2(1+) .
14 We now seek to write moments in terms of the curvatures, Fig. 15.5, introducing the stresses
from Eq. 15.14 into Eq. 15.1-a, and using Eq. 15.10 we derive the rst term
t t t
2 2 E 2 E
Mxx = xx zdz = (xx + yy )zdz = (zxx + zyy )zdz
t t 1 2 t 1 2
2 2 2
t
E 2
2 Et3
= ( xx + yy ) z dz = (xx + yy ) (15.15-a)
1 2 t 12(1 2 )
2
15 Following a similar procedure for the other two terms, we obtain the following moment-
curvature relation
Mxx Et3
1 0 xx
Myy = 1 0 yy
12(1 2 ) 1 (15.16)
Mxy 0 0 2xy
2
M D
Et3
16 The 12(1 2 ) term is referred to as the exural rigidity and is analogous to the exural
stiness EI of a beam (if the plate has unit width, and = 0, then EI = Et3 /12).
17This theory, is primarily applicable to thick plates in which shear deformations can not be
neglected.
18In this formulation, shear deformation is accounted for and is the extension of the Timo-
shenkos beam theory.
Kinematic: The kinematic relation of Eq. 15.12 still holds and is written in slightly modied
157
form
xx
0
x xx
yy = 0 y (15.17-a)
yy
2xy
y
x
Lf
w
xz
= x 1 0 (15.17-b)
yz
y 0 1 xx
yy
Ls
u
Constitutive: the constitutive relation of Eq. 15.16 must be augmented to account for the
relationships between xz and yz (which are no longer negligible in this case) with Vx
and Vy respectively. This is obtained by equating the internal work with the external one
xz xz dzdA = Vx xz dA (15.18)
A z A
xz
Recalling that xz = ,
and from Eq. 15.40, we obtain
2 t/2 ' z (2 2
xz 1 3 Vx
xz xz dzdA = dzdA = 1 2 dzdA
A z A z A t/2 2 t t
8 3 Vx 2 t 6 Vx2 6 Vx
= dA = dA = Vx dA(15.19-a)
A 15 2 t A 5 t A 5 t
6 Vx
Hence, comparing with Eq. 15.18 we conclude that xz = 5 t , and the nal constitutive
matrix becomes
Mxx 1 0
xx
K
Myy 1 0 0
yy
Mxy = 0 0 1
2xy
2
1 0
(15.20)
Vx
xz
0
Vy 0 1 yz
M D
Et3
where K = 12(1 2 ) and = 56 t
19 The dierential equation for the Mindlin plate will not be derived in this course.
21 Substituting
Z " #
1
x x
+ y 2 y x y
1
+ y 2 x xx
W = K xx yy 1 1 dA
A | {z }
y x
+ x 2 y y y
+ x 2 x
yy
T | {z }| {z }
Ke
f
| {z }
Internal, Bending
2 38 9
Z
x x
+
y y
y
x < w =
5t
w xx yy 4
1 0 5 xx dA
6 A | {z } y
: ;
0 1 yy
uT | x
{z }| {z }
Ke u
s
| {z }
Internal, Shear
8 9 8 9
Z < pz = Z < ps =
+ w xx yy 0 dA + w xx yy 0 dA (15.23-a)
A | {z }: 0 ; | {z }: 0 ;
uT | {z } uT | {z }
pv p
| {z }
External
+ uT pV dA + uT pd = 0 (15.24)
A
Kef = LTf Df Lf (15.25)
Kes = LTs Ds Ls (15.26)
Z
8 9
< pz = Z
8 9
< ps =
159
w xx yy 0 dA + w xx yy 0 dA = 0 (15.27-a)
A : ; : ;
0 0
15.2.2 Kirchho
22 This theory, is primarily applicable to thin plates in which shear deformations can be ne-
glected.
23In this formulation, shear deformations are neglected, and formulation is analogous to the
conventional Euler-Bernouilli beam theory.
Kinematic Relations Since shear deformations are neglected, xz = yz = 0 and thus the
last two relation of Eq. 15.10 reduce to
Equilibrium The equilibrium equation, as expressed by Eq. 15.7 still holds. If we were to
substitute the second and third relations into the rst one, then we would obtain the
following equilibrium relation in terms of the moments
d2 M
(Note the similarity with the corresponding equations for beam exure Vx = 0)
dx2
24 If we combine the kinematic with the constitutive relation, 15.30 and 15.31 respectively we
obtain
2 2
Mxx + y
Et 3
2
x2 2
2
* +
Myy = y2 + x w (15.33)
12(1 )
2 2
Mxy (1 )
x y
25 Finally, we substitute the equilibrium equation, 15.32, into the previous one,
4w 4w 4w pz
4
+2 2 2 + = Et3
(15.34)
x x y y 4
12(1 2 )
or
pz
4 w = Et3
(15.35)
12(1 2 )
15.2.2.3 Stresses
26 Combining the stress-strain relation of Eq. 15.14, with Eq. 15.29, and 15.33, the stresses
My
again we note the analogy with the exural stress expression in beams = I . Using the three
dimensional equilibrium equation, Eq. 3.57:
xx xy xz
+ + = 0
x y z
yx yy yz
+ + = 0 (15.38-a)
x y z
zx zy zz
+ + = 0
x y z
integrating the rst equation yields
z t/2
xx xy 12z Mxx Mxy
xz = + dz = + dz
t/2 x y z t3 x y
12 t/2
= 3 zVx dz (15.39-a)
t z
Finally
# 2 $ # 2 $
3Vx 2z 3Vy 2z
xz = 1 and yz = 1 (15.40)
2t t 2t t
Victor Saouma Finite Elements II; Solid Mechanics
Draft
15.2 Plate Theories
Thus, the shear stress distribution across the plate thickness is parabolic (though they tend to
1511
be very small compared to xy , and the peak shear stresses occur at the middle surface (z = 0)
where
3 Vx 3 Vy
xz |max = and yz |max = (15.41)
2 t 2 t
Finally, it can be shown that zz varies cubically.
27 Prior to the nite element discretization, we seek to obtain from the previously derived
where M = D is obtained from Eq. 15.31, and = Lu is obtained from Eq. 15.30.
29 Accounting for the external virtual work, we obtain
W = Wi + We = T T
u (L DL)udA + T
u pv dA + uT pd = 0 (15.43)
A
A
p
where pv contains the applied transverse loading pz , and p the edge loading ps .
30 Substituting for the LT DL (Eq. 15.30) terms we obtain
2
1 0
x22
2 2 Et3 1
Ke = LT DL = x
y
2 x
0 y
2 2 y 12(1 2 ) 1
2
2
0 0 2 x y
= K
2 2 2 2 2 2
+ 2 + 2(1 ) + 2 + 2 (15.44-a)
x2 x2 y x y x y y y 2 x
15.2.3 Summary
31 Table 15.1 summarizes some of the major equations governing plate bending, and contrasts
2 0 0
3
8 9 x 8 9
>
> xx >
> 6 0 0 7 2
>
> >
yy > 6 y 7 8 9 8 9 > 2 >
>
>
< >
= 6 7 < ux = < xx = >>
< x >
zz 6 0 0 z
7 2
=
=6
6
7
7 u y yy 2 w
>
> xy >
> 6 y x
0 7 : uz ; : ;=>> y >
> | {z
>
> >
> 7 2xy > >
> xz > 0 2 x ; u }
: ; 64 5 | {z } | {z } : y
Draft
yz z x
0 u | {z }
| {z } z y
Victor Saouma
L
" | {z }
L 3
8 9 2 0 " #8 9
< xx x 1 0 < = w
xx xz
= 6 7
yy 0 ; = x xx
: ;=4 y 5 yy yz 0 1 : ;
2xy y yy
| {z } | y x | {z }| {z } | {z } | {z }
{z }
Ls u
Lf
Equilibrium
8 xx
9 9
2
3>>
> >
>
> 8 9 2
38> Mxx > 8 9 8 9
x
0 0 y z
0 >
>
< yy >
>
= < bx = x
0 y
1 0 >
>
< Myy
>
>
= < 0 = < 0 =
6 7 zz 6 7
6
4 0 y
0 x
0 z
7 + by 6
4 0 y x
0 1 7 Mxy + 0 0
5> xy > 5> > : ;=: ;
> > > > 0
: bz ; = 0 Vx pz
0 0 z
0 x y
>
>
> xz >
>
> 0 0 0 x y
>
: >
;
| {z }: ; | {z } | {z Vy
yz
b
}| {z }
LT | {z } LT M
9
2
38> Mxx > 8 9 8 9
x
0 y
1 0 >
>
< Myy
>
>
= < 0 = < 0 =
6 7
6
4 0 y x
0 1 7
5> Mxy + 0 0
> Vx
>
> : pz
;=: 0
;
0 0 0 x y
>
: >
;
| {z }| Vy
{z }
LT = LT
f LT
s M
Constitutive
8 xx
9 2 1 0 0 0
38 9
> > > xx >
>
>
> yy
>
>
> 6 1 0 0 0 7 >
> >
>
> 8 9 2 38 9
>
< >
= 6 7>> >
= < Mxx = 1 0 < xx =
1 0 0 0
zz
< yy
6 7 zz 1 0 5
= 6 7 Myy =K4 yy
> xy > 6 0 0 0 0 0 7 xy > : ; 1 : ;
>
> >
> 4 5>>
> >
> Mxy 0 0 2
2xy
>
>
: yz >
>
; 0 0 0 0 0 >
>
: yz >
>
; | {z } | {z }
zx 0 0 0 0 0 zx
{z }|
M D
| {z } | {z } | {z }
D " 2 3
8 Mxx
9 2 1 0
38
xx
9
>
>
> >
>
> 1 0 5 7 >
>
> >
>
>
< Myy = 66 K4 0 7 < yy =
1
Mxy =6
6 0 0 2
7 2xy
>
> >
> 7
5>> >
>
>
: Vx >
; 4 1 0 >
: xz >
;
0
Vy 0 1 yz
| {z } | {z }| {z }
M D
E(1) Et3
; = 12
= (1+)(12) ; = 1 2(1)
K = ; = 5
6
t
12(1 2 )
Virtual Work Z Z
R R R
= (Lu)T Dd uT bd uT td = 0 W = w (KLwpz ) dA + p wps d = 0
t A
Z 6ks (1 )
Z Z Z
T e T e T T
W = K Kf dA + u Ks udA + u pV dA + u pd = 0
A t2 A A
32 Numerous elements have been proposed for plate bending. An ideal plate element should
33 Since plate structure is a special case of shell structures, and plates under large displacement
exhibit shell characteristics (through membrane actions), similar formulations can be employed
for both plates and shells.
34 The stiness matrix, including shear deformation eects, for a rectangular element, Fig. ??
will be developed.
Note that the natural coordinate system selected has its origin at node 1, and 0 , 1. Thus
PLATES
through these shape functions, we will have a rst order (C0 ) element, for which displacement
and rotations will be continuous along the boundaries (since xx and yy are also variables).
36 The stiness matrix, for a Mindlin plate, will be derived on the basis of the principle of
t 2
39 We note that Lf = Lu (used in elasticity) for in plane deformation, and Df = 12 D, thus
the stiness matrix for bending can be formed directly from the stiness matrix for in-plane
deformation.
40 Upon substitution and integration we obtain
t2
Kef = Et
24(1 2 ) 12
2 w1 w2 w3 w4 x1 x2 x3 x4 y1 y2 y3 y4 3
6 0 0 0 7
6 0 7
6 7
6 0
A /2 3 2 7
6 A C B 2 3 7
6 A B A /2 3 2 3 2 7
6 7
6 A C 2 3 2 3 7
6 3 2 7
6 A 2 3 7
6 A B A /2 C 7
6 A /2 7
6 A C 7
6 A B 7
4 A
5
(15.52)
Note that the rst row and second row in the second equations correspond to xz and yz
respectively.
42 Upon substitution,
1 2
B11 = (1)
a
(1)
a
a a (15.56-a)
! "
B12 = (1 )(1 ) (1 ) 1 (15.56-b)
1 2
B21 = (1)
b
b
b
(1)
b
(15.56-c)
! "
B12 = (1 )(1 ) (1 ) (1 ) (15.56-d)
which gives
PLATES
1
Ks =
24
w1 w2 w3 w4 x1 x2 x3 x4 y1 y2 y3 y4
2 3
6 6 6 6 3b 3b 3b 3b 3a 3a 3a 3a
6 6 6 6 3b 3b 3b 3b 3a 3a 3a 3a 7
6 7
6 6 6 3b 3b 3b 3b 3a 3a 3a 3a 7
6 7
6 6 3b 3b 3b 3b 3a 3a 3a 3a 7
6 7
6 1.5ab 1.5ab 1.5ab 1.5ab 7
6 7
6 1.5ab 1.5ab 1.5ab 7
6 7
6 7
6 1.5ab 1.5ab 7
6 7
6 1.5ab 7
6 1.5ab 1.5ab 1.5ab 1.5ab 7
6 7
6 1.5ab 1.5ab 1.5ab 7
6 7
4 1.5ab 1.5ab 5
1.5ab
(15.59-a)
where = ab , = ab , = + , and = .
44 The nal element stiness matrix is
! "
Ke = K Kef + Kes (15.60)
P
[Kf + Ks ] u = (15.61)
K
45 Through inspection of the previous equation, and noting that = 6ks (1)
t2
, we observe that
for very thin plates is very large, hence as t 0, . Hence, unrealistic u = 0 can be
obtained independently of the load. This phenomena is called locking (of the displacements).
46 Locking occurs also in thin beam elements when shear deformation eects are accounted for.
47 To alleviate the locking problems is to make Ks singular so that Ks is nite. This can
be accomplished through reduced numerical integration for Ks , Table 15.2, which in turn will
reduce the rank of the matrix.
Integration Rule
Full Reduced Selective
Kf Ks Kef Kes Kef Kes
e e
48 The same phenomena occurs for beams. This is illustrated by evaluating the deection of a
1517
cantilever beam, due to both exural and shear deformation, under a point load
P L3 PL P L3 3EI
w= + = 1+ (15.62)
3EI ks A 3EI ks AL2
or 2 2
P L3 3 ks AL2 L L
w= 1+ ; = = ks =O (15.63)
3EI EI E r h2
Hence, as the slenderness ratio increases, shear deformation decreases. Furthermore, Euler
theory is recovered if the shear stiness , however shear locking can occur for small .
49 Shear locking eects are aggravated by large element distortions.
50 Since rectangular elements can hardly accommodate general boundary conditions, and since
quadrilateral elements can more readily be obtained through an assemblage of triangular ele-
ments rather than through the distortions of a rectangular element, we consider next the three
PLATES
slopes xx = w
x and yy = y . Because the complete cubic polynomial expansion contains
w
10 terms (1, , , 2 , , 2 , 3 , 2 , 2 , 3 ), yet we have only nine dofs, thus we opt instead to
a natural coordinate system and we express the displacement as
w = w1 + w2 (15.64)
where w1 corresponds to the rigid body part, and w2 the displacements incurred when the
triangle is simply supported at the nodes.
52 w1 has three rigid body motions, Fig. 15.7
2. Rotation about side 1: of the triangle. This results in a rigid body translation in the
z direction of line ab which is parallel to side 1 is w1 = h1 . If we consider a triangle
with vertex b, than this point would have L1 = A1 /A = h1 /h, thus w1 = h1 = hL1 or
w1 = (cst)L1 .
1 L1 + w
w1 = w 2 L2 + w
3 L3 (15.65)
53 Since the element has 9 dof, a cubic polynomial will be used for shape functions; i.e. the
56 Hence
1519
w = w1 + w2 = Nu w (15.69)
are dened in Eq. 15.65 and 15.68-a.
where Nu and w
57 The unknown quantities w i can in turn be expressed in terms of the nodal unknown dis-
placements by substituting w, xx = w
x and yy = y in Eq. 15.69.
w
w = Nu Gu = Nu (15.71)
where
Ke = LT DL (15.74)
Thus
Ke = NT LT DLNdA (15.75)
A
e
p = NT pz dA (15.76)
A
15.3.2.2 Nonconformity
1
PLATES
y
1
2 3
x 2 3
2
62 However, for yy = w
y things are quite dierent. If we set L1 = 0 and take the derivative
of w in Eq. 15.69, then we have
w
yy = = 0 + 1 x + 2 x2 (15.78)
y
but since we only have two d.o.f. (yy at nodes 2 and 3), then the parameters can not be
uniquely expressed in terms of these two d.o.f. Hence, we will have discontinuity of normal
slope, or kinks.
63 Elements with element boundary discontinuity are referred to as nonconforming elements.
64 This nonconforming element, nevertheless is a viable one as the 3 rigid body modes, and
5
y
1
4
2
x
69 Since for bending, only the rotations appear in the formulation, the rotations can be approx-
imated by
6
6
xx = Ni xi ; yy = Ni yi (15.80)
i=1 i=1
70 The shape functions Ni for the natural coordinates with quadratic expansions are
71 For this element, so far we have a total of 12 dof, and we seek to formulate it in terms of only
i i
9 dof, 3 at each corner node corresponding to wi , xx and yy thus we need to impose certain
constraints for this reduction.
72 Hence, Kirchho assumptions will be selectively applied:
w
xz 1 0 0 xx w,x
= x = = (15.82)
yz
y 0 1 xx 0 yy w,y
yy
tk = w,s
k
(15.83)
where k is a superscript denoting the midside nodes only, tk is a rotation about the normal
to the boundary at node k, and wsk is the derivative of w with respect to the s direction
along the boundary.
3. Linear variation of the tangential slope: , i.e the rotation about the s direction at the
middle nodes is the average of the rotations at the end nodes
1
nk = (ni + nj ) (15.84)
2
where the corner nodes i j are 2-3, 3-1, and 1-2. k corresponds to the midside nodes.
and thus w = w,s varies quadratically and the three coecients can be obtained by matching
w,s = n at the three points along the edge. Applying the boundary conditions at s = 0 and
s = lij in terms of both w and w,j , we have 4 equations to determine the four coecients,
yielding
3 i 1 i 3 j 1 j
k
w,s = w w,s + w w,s (15.87)
2lij 4 2lij 4
74 Since the rst term of Eq. 15.22-c, does not involve w, there is no need to dene an interpo-
and
w,n cos ij sin ij xx
= (15.89)
w,s sin ij cos ij yy
76 Substituting the last three equation into Eq. 15.80 at each node to obtain xi and yi to nd
1523
xx = Nx vi ; yy = Ny vi (15.90)
or
w1
1
x
1
y
i
w2
xx Nx1 Nx2 Nx3 Nx4 Nx5 Nx6 Nx7 Nx8 Nx9 2
= x (15.91)
i
yy Ny1 Ny2 Ny3 Ny4 Ny5 Ny6 Ny7 Ny8 Ny9
2
y
N
w3
3
x
3
y
v
where
Nx1 = 1.5(a6 N6 a5 N5 ), Ny1 = 1.5(d6 N6 d5 N5 )
Nx2 = N1 b5 N5 b6 N6 , Ny2 = Nx3
Nx3 = c5 N5 c6 N6 , Ny3 = N1 e5 N5 e6 N6
Nx4 = 1.5(a6 N4 a5 N6 ), Ny4 = 1.5(d4 N4 d6 N6 )
Nx5 = N2 b5 N6 b6 N4 , Ny5 = Ny3 (15.92)
Nx6 = c6 N6 c4 N4 , Ny6 = N2 e5 N6 e4 N4
Nx1 = 1.5(a6 N5 a5 N4 ), Ny1 = 1.5(d5 N5 d4 N4 )
Nx2 = N3 b5 N4 b5 N5 , Ny2 = Nx3
Nx3 = c4 N4 c5 N5 , Ny3 = N3 e4 N4 e5 N5
x x 1 y2
1 2
3 xij yij y y 1 x2
1 2
and ak = l2ij , bk = 4 ij 2 ij
2
lij
, ck = 4 lij2 , dk = l2ij , ek = 4 ij 2 ij
2
lij
and k = 4, 5, 6
ij ij
corresponds to sides ij=23, 31, and 12 respectively.
77With the shape functions thus dened, we can nally determine the stiness matrix from
the principle of virtual work (Eq. 15.22-c). Since shear deformations are neglected, we use only
the rst part.
Wi = T LTf Df Lf dA (15.93-a)
A
Kef
T
= vi NT LTf Df Lf NdAvi (15.93-b)
A
T
= vi BT Df BdA vi (15.93-c)
A
Ke
(15.93-d)
15.4 Summary
PLATES
78 Major conclusions
4. M , .
5. If thick plate theory is applied to thin plates, will have shear locking problems.
7. For thin plate theory, will have non-conforming elements if only 3 d.o.f. are adopted at
each node.
10. Shear locking is primarily an issue for 3 node and 4 node elements. Higher order ex-
pansions perform much better, whereby Lagrangian 6 and 16 node elements perform
signicantly better than their Serendipity counterparts of 8 node and 12 node elements.
11. The usual 3 DOF/node Kirchho elements do not maintain normal slope continuity.
Chapter 16
MATERIAL NONLINEARITIES
16.1 Introduction
16.1.1 Linearization
= () (16.1)
Neglecting quadratic and higher order terms leads to a linearized constitutive law
() + D() (16.3)
D (16.4)
or
MATERIAL NONLINEARITIES
int ext
f (u) = f (16.6)
5 We now develop a linearized expression for the internal forces. Given u as nodal displacements
yielding strain eld = Bu, and the stress eld = () = (Bu). Then, the Taylor expansion
of the internal forces around u yields
f
int
int int
f (u + u) = f (u) + u + (16.7)
u
u=u
where
int
f
KT (16.9)
u
which is the well known formula for the stiness matrix, however Del is now replaced by the
tangent moduli D
7 Before we discuss solution strategies, it may be helpful to point out the parallelism which
exists between (numerical) solution strategies, and (experimental) testing methods. Modern
testing equipment can be programmed to apply a pre-determined rate of load (as measured by
a load cell), of displacement (as measured by an internal displacement transducer), or of strain
(or relative displacement such as crack mouth opening displacement) measured by a strain/clip
gage or other instruments, Fig. 16.1
Load Control: the cross-head applies an increasing load irrespective of the specimen deforma-
tion or response. For all materials, when the tensile strength is reached, there is a sudden
and abrupt brittle failure. The strain energy accumulated in the specimen is suddenly
released once the ultimate load of the specimen is reached, thus the sudden failure can be
explosive.
P
163
11
00 11
00 11
00 111
000
11
00
P, , CMOD
, 11
00
11
00 11
00 11
00 111
000
11
00 11
00 11
00 111
000
Actual
Programmed
CMOD
Strain Control: is analogous to displacement control, except that the feedback is provided by
(strategically positioned) strain gage or a clip gage or an arbitrary specimen deformation
(not necessarily corresponding to the loading direction). To accomplish this test a clip
gage or a strain gage has to provide the feedback signal to the testing equipment in order
to accordingly adjust the stroke.
8 Similarly, the objective of a nonlinear nite element analysis is to trace the (nonlinear)
response of a structure subjected a given load history. This is best done in an incremental-
iterative procedure where the load (or the displacement) is applied through several increments,
and within each increment we seek to satisfy equilibrium through an iterative procedure (caused
by the nonlinearity of the problem).
9 The incremental analysis can be performed under
1. Load control; Load is incrementally applied on the structure.
2. Direct displacement control; An imposed displacement is applied.
3. Indirect displacement control (such as relative displacements between two degrees of free-
dom)
4. Arc-Length control
16.2.1 Newton-Raphson
12For the sake of discussion, we will assume in the following sections that the incremental
ext
analysis is under load control, with increments of loads f .
13 At the end of each load increment, internal forces must be in equilibrium with the external
u = un , Rn = 0 (16.12)
ext
and apply an increment of load f n such that
ext ext ext
f n+1 = f n + f n (16.13)
un+1 = un + un (16.14)
ext
We will keep f n reasonably small to capture the full nonlinear response.
16This is the most rapidly convergent process (albeit computationally expensive) of non-linear
problems.
17 At the beginning of each step n + 1, we start from the displacement un that were computed
int ext
in the previous step through equilibrium Rn 0 or f n f n . The external forces are
ext ext ext ext
now increased from f n to f n+1 = f n+1 + f , and we seek to determine the corresponding
int ext
displacements un+1 through equilibrium Rn+1 0 or f n+1 f n+1 .
18 Within the current step (identied through the subscript n), we will be iterating (through
f ext
165
2 3
Rn+1 Rn+1
f ext
n+1 n+1
f int,
2
n+1
fnext
Rn+1
1
1 f int, n+1
K 1
T
n
f ext
n un1 un2
un1
un2
un 1
un+1 2
un+1 3
un+1 u
or
very rapidly.
23 It should be noted that each iteration involves three computationally expensive steps:
24 This method is essentially the same as the Newton-Raphson however in Eq. 16.23 (KiT ) is
replaced by KT which is the tangent stiness matrix of the rst iteration of either 1) the rst
increment KT = K1T,0 , Fig. 16.4, or 2) current increment, Fig. 16.3 KT = K1T,n Fig. 16.3
f ext
2 3 4
Rn+1 Rn+1 Rn+1
f ext
n+1 n+1
1
K
T
f int,
3
n+1
f int, n+1
fnext
2
Rn+1
1
1 f int, n+1
K 1
T
n
f ext un1 un2 un3
n
un1
un2
un3
un 1
un+1 2
un+1 3
un+1 u
25In general the cpu time required for the extra iterations required by this method is less than
the one saved by the assembly and decomposition of the stiness matrix for each iteration.
26 It should be mentioned that the tangent stiness matrix does not necessarily have to be the
true tangent stiness matrix; an approximation of the true tangent stiness matrix or even the
initial stiness matrix will generally produce satisfactory results, albeit at the cost of additional
iterations.
27 This method is a compromise between the rst two. First we seek two displacements by two
cycles of modied Newton-Raphson, then a secant to the curve is established between those
f ext
167
2 3 4
Rn+1 Rn+1 Rn+1
f ext
n+1 n+1
0
K
T
0 f int, n+1
fnext
K 3
Rn+1
T
1
f int,
2
n+1
f int,
1
n+1
n
f ext un1 un2 un3
n
un1
un2
un3
un 1
un+1 2
un+1 3
un+1 u
28 Subsequently, each step will be taken along a secant connecting the previous two points.
f ext
MATERIAL NONLINEARITIES
2 3 4
Rn+1 Rn+1 Rn+1
f ext
n+1 n+1
f int,
3
n+1
f int,
2
n+1
K0
fnext
Rn+1
T
1
f int,
1
n+1
n
f ext un1 un2 un3
n
un1
un2
un3
un 1
un+1 2
un+1 3
un+1 u
33For k = 0 and k = 1, the values of sk are 0.0 and 1.0, respectively. Therefore, uj+1,0 = uj
and uj+1,1 = uj+1 . The orthogonality condition is quantied by a scalar value gk representing
the iterative change in energy, which is dened as
j+1,k
gk = uj R (16.30)
where
j+1,k ext int
R =f f (uj+1,k ) (16.31)
are the residual loads at the end of solution iteration j and line search iteration k.
34 gk can be expressed as a function of sk (see Figure 16.6) and the object of the line search is
to nd sk such that gk is zero. An estimate of sk+1 such that gk+1 is zero can be computed
using a simple extrapolation procedure based on similar triangles
sk+1 sk
= (16.32)
g0 g0 gk
On rearranging terms, sk+1 is dened as
k+1 k g0
s = s (16.33)
g0 gk
35 As a preventative measure, sk+1 is assigned a value of 5.0 for all sk+1 > 5.0 so that unre-
169
int
strained over-relaxation is inhibited. Once sk+1 is estimated, uj+1,k+1 , f j+1,k+1 , and Rj+1,k+1
are computed for the next line search iteration, Fig. 16.6.
g
g0
g1
s1 s2 s
and g0 gk 0.001 | g0 |. Smaller tolerances may be used to determine if the line search has
converged, Zienkiewicz and Taylor (1989) prefer to use 0.6, but Criseld, M.A. (1979) concluded
that there was little advantange to be gained by doing such.
37 The owchart illustrating the Line Search algorithm is shown in Fig. 16.7.
Initialize s1
0
Compute R
DO 30 k=1,3
Do 10 j=0, Niter
Compute gk and sk+1
j j+1 0
Compute u, u, g
j+1,k+1 j+1,k+1
Compute and
Yes
Converged j+1,k+1
j+1,k+1
Compute fint and R
No
10
30
38 In all preceding methods, iterations are performed until one or all of a variety of convergence
criteria are satised. Relative convergence criteria are optionally enforced on the displacements,
loads, and/or incremental energy to dene the termination conditions.
39The relative displacement criteria is dened in terms of the displacement corrections uj and
the updated incremental displacements uj+1 as
uj 2
Ou = (16.35)
uj+1 2
v 2 = vi2 (16.36)
i=1
43 Independently of the choice of iterative algorithm, any solution strategy using load control
1611
fails if the prescribed external loads cannot be maintained in equilibrium by the internal forces.
This would typically occur if the load is monotonically increased until the load-carrying capacity
of the structure is exhausted, Fig. 16.8
f ext
44 In most engineering analyses, it is simply required to determine the maximum load carrying
capacity, and the corresponding displacements. As such, divergence of the iterative process is
often taken as an indicator of structural failure, and the last converged step provides information
on the state prior to collapse.
45 However, nite element simulations of complex engineering problems can diverge for a num-
ber of other reasons, many of which are purely numerical and have nothing to do with the real
structural failure.
46 If the load-displacement diagram is to be followed beyond the peak, i.e post-peak response
is required, then alternative solution strategy to the load-control one must be devised.
47 Post-peak response may be of interest not only in problems in structures with imposed
displacements (such as initial settlements), but also to assess the ductility of the structure
(specially when cracks are present).
48 To outline the displacement controlled algorithm, we divide the displacements into two
groups: one with unknown displacements at nodes that are left free, and the second with
prescribed displacements at nodes that are controlled. Accordingly, we partition the displace-
ment vector into {uf , up }T and the internal and external force vectors into {f int,f , f int,p }T and
{f ext,f , f ext,p }T , respectively. External forces f ext,f (corresponding to the unknown displace-
ments uf ) are prescribed, and for simplicity we will assume that they are equal to zero. All
external forces acting on the structure are represented by reactions f ext,p at the supports with
prescribed displacements up .
49 Hence, the equilibrium equations are partitioned as
f int,f (uf , up ) = 0 (16.41-a)
f int,p (uf , up ) = f ext,p (16.41-b)
50 For given up , the unknown displacements uf can be computed by solving Eq. 16.41-a. After
that, the reactions f ext,p are obtained by simple evaluation of the left-hand side in (16.41-b).
51
MATERIAL NONLINEARITIES
(n1)
and up from the previous step, and we replace Eq. 16.41-a by the linearized equations
(n1) (n1) (n,1) (n1)
ff + K11 uf + K12 u(n,1)
p =0 (16.42)
f int,f f int,f
where K11 uf and K12 up are blocks of the global tangent stiness matrix
f int,f f int,f
f int uf up K K
K = = 11 12
(16.43)
u f int,p f int,p K21 K22
uf up
where
(n,0) (n1)
uf = uf (16.44-a)
u(n,0)
p = u(n1)
p (16.44-b)
u(n,1) p up
= u(n) (n1)
p (16.44-c)
u(n,i)
p = 0 for i = 2, 3, . . . (16.44-d)
55 Note that, starting from the second iteration, the correction up is zero, and so the term with
K12 on the right-hand side of (16.3) vanishes. This term is present only in the rst iteration. It
(n,0) (n) (n,0) (n1)
might seem that one could start immediately from up = up instead of up = up , and
then the correction up would be zero already in the rst iteration and the matrix K12 would
never have to be evaluated. However, this is in general not a good idea because such an initial
approximation would be too far from the equilibrium path and the process might diverge.
56 Direct displacement control can be applied only on structures loaded only at one point, or
when the load is transmitted by a sti platen so that all points on the loaded surface exhibit
the same displacements.
57 However, this is not always the case. As an example, consider a dam loaded by hydrostatic
pressure due to reservoir overow; see Fig. 16.9. Here, the load is applied along a large portion of
the boundary, and the shape of the corresponding displacement prole is not known in advance.
Another case in which direct displacement control fails is very brittle failure characterized by
a load-displacement diagram with a snapback, Fig. 16.10.
u u u
58 Advanced incrementation control techniques abandon the assumption that the values of
external loads and/or displacements at supports after each incremental step are prescribed in
advance. Instead, the loading program is parameterized by a scalar load multiplier.
59
MATERIAL NONLINEARITIES
Restricting the applied loading to be proportional, a scalar load parameter can be used to
ext
scale an arbitrary set of applied loads f .
ext
60 The applied loads at the start of increment i are dened as the scalar-vector product i f ,
where i is the load parameter at the start of increment i. i is zero at the start of the rst
increment.
61The applied incremental loads for increment i are dened as the scalar-vector product
ext
i f , where i is the incremental load parameter for increment i. The updated load
parameter i+1 at the end of increment i is
i+1 = i + i (16.45)
62 The incremental displacements due to the applied incremental loads are obtained using the
standard modied-Newton algorithm, as described in Zienkiewicz & Taylor (1991). The in-
cremental displacements uj+1 at the end of iteration j for a generic increment are dened
as
Duj+1 = uj + uj (16.46)
where uj are the incremental displacements at the start of iteration j and uj are the incre-
mental displacement corrections for iteration j.
63The incremental load parameter j+1 at the end of iteration j is dened in an analogous
manner as
j+1 = j + j (16.47)
where j is the incremental load parameter at the start of iteration j and j is the incremental
load parameter correction for iteration j. At the start of the rst iteration uj and j are
identically zero.
64 Incremental displacement corrections are determined by solving
ext ext ext j
K uj = (f + j f + j f f int ) (16.48)
elem
j
f int = BT D ( + j ) (16.49)
e=1 e
are the reactions for the state of stress at the start of iteration j.
j
65 Dening the residual forces R at the start of iteration j as
j ext ext j
R = f + j f f int (16.50)
ext
66 The matrix-vector product K1 f is invariant for the increment and, therefore, can be
treated as a vector constant uT , which Criseld (1981) referred to as the tangent displacements
uT = K1 f
ext
(16.52)
67
j
The matrix-vector product K1 R denes the displacement corrections ujr due to the resid-
1615
ual forces
ujr = K1 R
j
(16.53)
but they are obviously not invariant for the increment. The displacement corrections for itera-
tion j are then dened as
uj = j uT + ujr (16.54)
68 Figure 16.11 shows a owchart for an incremental nonlinear nite element program based on
the modied-Newton algorithm with indirect displacement control capabilities. The numbers
in the boxes in Figure 16.11 correspond to those appearing in Figure ??.
16.4.2 Arc-Length
Adapted from (Jirasek and Bazant 2001)
69 The basic idea of a exible incrementation control technique is that the step size is specied
by a constraint equation that involves the unknown displacements as well as the load multiplier.
The original motivation was provided by the requirement that the size of the step measured as
the geometric distance between the initial and nal state in the load-displacement space should
be equal to a prescribed constant, Fig. 16.10.
70 Despite the apparent simplicity of the condition of a constant arc length, it must be used with
caution. First of all, it is important to realize that forces and displacements have completely
dierent units, and so the purely geometrical measure of length in the load-displacement space
does not make a good sense. It is necessary to introduce at least one scaling factor, denoted as c,
that multiplies the load parameter and converts it into a quantity with the physical dimension
of displacement. The length of a step during which the load parameter changes by and the
displacements change by u is then dened as
5
l = uT u + (c )2 (16.55)
71By adjusting the scaling factor we can amplify or suppress the relative contribution of loads
and displacements. One reasonable choice is derived from the condition that the0 contributions
should be equal as long as the response remains linear elastic, which leads to c = uTe ue where
ue is the solution of Ke ue = f .
72 In some cases, e.g., for frame, plate, and shell models that use both translational and rota-
tional degrees of freedom, the components of the generalized displacement vector u do not have
the same physical dimension. It is then necessary to apply scaling also to the vector u.
73 Consider an incremental solution process controled by the arc-length method. In a typical
step number n, we start with displacements u(n1) and load parameter (n1) computed in
the previous step, and we search for displacements u(n) and load parameter (n) . The state
at the end of the step must satisfy the equations of equilibrium between the internal forces
f int (u(n) ) and external forces f ext ((n) ). Compared to the load control or direct displacement
control, the load parameter is an additional unknown. The corresponding additional equation
is provided by the constraint that xes the size of the step. For example, we can require that
the length of the step evaluated from formula (16.55) be equal to a prescribed value, l. We
could treat the problem as a system of Ndf + 1 nonlinear equations, where Ndf is the number
Figure 16.11: Flowchart for an incremental nonlinear nite element program with indirect
displacement control
Figure 16.12: Two points on the load-displacement curve satisfying the arc-length constraint
where u(n,i) is the unknown displacement correction, and (n,i) is the unknown correction of
the load parameter. The rst three terms on the right-hand side are known, and the last term
is an unknown scalar multiple of a given vector f . We can therefore separately solve equations
(n,i1)
K(n,i1) u0 = f 0 + (n,i1) f f int (16.57-a)
(n,i1)
K uf = f (16.57-b)
we obtain a quadratic equation for a single unknown, (n,i) . This equation usually has two real
roots, corresponding to the two points of the equilibrium path that have the prescribed distance
from point (u(n1) , (n1) ); see Fig. 16.12. The correct root is selected depending on the sense in
which we march on the equilibrium path (Criseld, M.A. 1981), and the displacement correction
is determined from (16.58). After standard updates of the displacement vector and the load
parameter, the iteration cycle is repeated until the convergence criteria are satised.
participates in the failure mechanism. In cases when the failure pattern is highly localized,
robustness of the technique may deteriorate. The remedy is to adapt the constraint equation
to the particular problem and control the incrementation process by a few carefully selected
displacement components.
75 Motivation is again provided by the physical background. If the load-displacement diagram
the point at which the load-displacement diagram snaps back. The same idea can be exploited
by a numerical simulation. It suces to select a suitable linear combination of displacement
components that increases monotonically during the entire failure process, and to use this
combination as the control variable.
76 de Borst (1985,1986) concluded that arc-length methods (Riks 1979, Ramm 1981, Criseld
1981), which were the original IDC methods, were not satisfactory for analyses involving crack-
ing accompanied by softening.
77 The main problem with the arc-length methods, when used in this context, was that the
constraint involved all displacement components equally when, in fact, only a few displacement
components were dominant. The dominant displacement components were typically those for
nodes at or near the crack mouth. This being the case, de Borst proposed using a transformed
relative displacement component between two nodes as the constraint.
78 The transformed relative displacement component can dene the crack mouth opening dis-
placement (CM OD), crack mouth sliding displacement (CM SD), or some arbitrary displace-
ment u between two points on a structure. The arbitrary displacement u may correspond
to a relative displacement measured during an experiment such as the relative vertical displace-
ment between a point on the neutral axis of a 3-point bend beam over a support and the bottom
of the beam at mid-span.
79 As it is the most general case, the relative displacement criterion will be described in terms
of the arbitrary relative displacement u. A pair of nodes, m and n, are selected to dene u,
with their total displacements being (u)m and (u)n , respectively. The direction associated with
u is dened by a unit vector v. u is thereby dened as
82 Employing a procedure proposed by Criseld (1983) for use with the arc-length method,
the convergence of the solution alogrithm can be accelerated by performing approximate line
searches; approximate line searches under xed (i.e. non-scalable) loads are described in Section
??.
83 This procedure requires an extra iterative loop at the beginning of the line search loop in
which a combination of j and sk+1 satisifying the constraint conditions (i.e. Equations ??
and 16.60) is computed. As j is initially computed for s1 = 1.0, any change in sk requires
a corresponding change in j for the IDC constraint to remain satised. Consequently, an
iterative loop, in which j is recomputed based on the estimated value of sk+1 , is required to
obtain a compatible combination of j and sk+1 .
84 After recomputing j , the values of g0 and gk are also recomputed using Equation 16.30
j+1,k j+1,k
to reect the change in the residual loads R caused by the new value of j . f int is not
j+1,k
updated to reect the changes in sk+1 when recomputing R , which is strictly not correct,
but it does signicantly reduce the number of computations without causing any diculties
(Criseld 1983).
85 Finally, from the new values of g0 and gk , sk+1 is re-estimated using Equation 16.33. The
loop is terminated when
new s
| sk+1 k+1 |
0.05 (16.63)
| sk+1
new |
which generally requires only a few iterations. A ow chart of this procedure is shown in Figure
16.13.
86 Since the total displacements uj+1,k are now dened as
reecting the introduction of the relaxation parameter sk , the IDC constraint equations must
be modied accordingly. j for the stress criterion is now dened as
1 2
ft (j )n + sk (jr )n (n)n
j = min (16.65)
sk (T )n (n)n
It is these general forms of the constraint equations that are implemented in MERLIN.
Figure 16.13: Flow chart for line search with IDC methods
Bibliography
Abramowitz, M. and Stegun, I.: 1970, Handbook of mathematical functions, Technical report,
National Bureau of Standard. Applied Mathematics Series, No. 55.
Babuska, I.: 1971, Error-bounds for nite element methods, Numer. Math 20(3), 179192.
Babuska, I.: 1973, The nite element method with lagrange multipliers, Numer. Math.
20(3), 179182.
Brezzi, F.: 1974, On the existence uniqueness and approximation of saddle-point problems
arising from lagrangian multipliers, RAIRO 8-R2 pp. 129151.
Cervenka, V., Keating, S. and Felippa, C.: 1993, A comparison of strain recovery techniques for
the mixed iterative method, Communications in Applied Numerical Methods 9, 925932.
Criseld, M.A.: 1979, A faster modied newton-raphson iteration, Computer Methods in Ap-
plied Mechanics and Engineering 20(3), 267278.
Criseld, M.A.: 1981, A fast incremental/iterative solution procedure that handles snap
through, Computers and Structures 13(1-3), 5562.
Delaunay, B.: 1934, Sur la Sphere Vide, Izv. Akad. Nauk SSSR, Otdelenie Matematicheskii i
Estestvennyka Nauk 7, 793800.
Eisenberg, M. A. and Malvern, L.: 1973, On nite element integration in natural co-ordinates,
International Journal for Numerical Methods in Engineering 7, 574575.
Felippa, C.: 1999, Finite element lecture notes, Technical report, Dept. of Aerospace Engineer-
ing, University of Colorado, Boulder.
Felippa, C.: 2000, Lecture notes in advanced nite element methods (asen 5367),
Technical report, Dept. of Aerospace Engineering, University of Colorado, Boulder.
http://caswww.colorado.edu/courses.d/AFEM.d/Home.html.
Haser, N. and Sullivan, J.: 1991, Real Analysis, Dover Publications, New-York.
Jirasek, M. and Bazant, Z.: 2001, Inelastic Analysis of Structures, John Wiley, Chichester.
Okabe, A., Boots, B., Sugihara, K. and Chiu, S.: 2000, Spatial Tessellations; Concepts and
Applications of Voronoi Diagrams, John Wiley & Sons.
Draft
2 BIBLIOGRAPHY
Ottosen, N. and Petersson, H.: 1992, Introduction to the Finite Element Method, Prentice-Hall.
Reich, R.: 1993, On the Marriage of Mixed Finite Element Methods and Fracture Mechanics:
An Application to Concrete Dams, PhD thesis, University of Colorado, Boulder.
Schey, H.: 1973, Div Grad Curl and all That; An Informal Text on Vector Calculus, W.W.
Norton.
Terzaghi, K. and Peck, R.: 1967, Soil Mechanics in Engineering Practice, 2nd edition, John
Wiley & Sons, New York, NY.
Cervenka, J.: 1994, Discrete Crack Modeling in Concrete Structures, PhD thesis, University of
Colorado, Boulder.
Voronoi, G.: 1907, Nouvelles Applications des Parametres Continus a la Theorie des Formes
Quadratiques, J. Reine Angew. Math 133, 97178.
Xue, W. and Atluri, N.: 1985, Existance and stability, and discrete bb and rank conditions
for general mixed-hybrid nite elements in elasticity, Hybrid and Mixed Finite Element
Methods pp. 91112.
Zienkiewicz, O. C. and Taylor, R. L.: 1989, The Finite Element Method, Vol. 1, Basic Formu-
lation and Linear Problems, 4th ed., McGraw-Hill, London.
Zienkiewicz, O. C., Vilotte, J. P., Toyoshima, S. and Nakazawa, S.: 1985, Iterative method for
constrained and mixed approximations. an inexpensive improvement of fem performance,
Computer Methods in Applied Mechanics and Engineering 51(13), 329.
Appendix A
VECTOR OPERATIONS
1 This appendix covers some elements of Vector Calculus essential to understand some of the
2A eld is a function dened over a continuous region. This includes, Scalar Field g(x),
Vector Field v(x), or Tensor Field T(x).
3 We rst introduce the dierential vector operator Nabla denoted by
i+ j+ k (A.1)
x y z
4 We also note that there are as many ways to dierentiate a vector eld as there are ways of
5 The derivative of a vector p(u) with respect to a scalar u, Fig. A.1 is dened by
dp p(u + u) p(u)
lim (A.2)
du u0 u
u)
( u + p (u)
p
N T
dp
= T (A.4)
ds
dT
= N (A.5)
ds
B = TN (A.6)
curvature (A.7)
1
= Radius of Curvature (A.8)
we also note that p dp
ds = 0 if dp
ds = 0.
dp d ! 2 "
= (t + 1)i + (4t 3)j + (2t2 6t)k = 2ti + 4j + (4t 6)k (1.9-a)
dt dt
dp 0
= (2t)2 + (4)2 + (4t 6)2 (1.9-b)
dt
2ti + 4j + (4t 6)k
T = 0 (1.9-c)
(2t)2 + (4)2 + (4t 6)2
4i + 4j + 2k 2 2 1
= 0 = i + j + k for t = 2 (1.9-d)
(4)2 + (4)2 + (2)2 3 3 3
A.1.2 Divergence
9 The divergence of a vector eld of a body B with boundary , Fig. A.3 is dened by
considering that each point of the surface has a normal n, and that the body is surrounded by
a vector eld v(x). The volume of the body is v(B).
v(x)
n
where v.n is often referred as the ux and represents the total volume of uid that passes
through dA in unit time, Fig. A.4 This volume is then equal to the base of the cylinder dA
times the height of the cylinder vn. We note that the streamlines which are tangent to the
boundary do not let any uid out, while those normal to it let it out most eciently.
11 The divergence thus measure the rate of change of a vector eld.
n
VECTOR OPERATIONS
dA
v
v.n
12 The denition is clearly independent of the shape of the solid region, however we can gain
an insight into the divergence by considering a rectangular parallelepiped with sides x1 , x2 ,
and x3 , and with normal vectors pointing in the directions of the coordinate axies, Fig. A.5.
If we also consider the corner closest to the origin as located at x, then the contribution (from
x3
e3 -e 1
x3
-e x1
2 e2
x2
x2
e1
-e 3
x1
13 or alternatively
A5
2 F F = F,ii (1.16)
v = i+ j+ k (x2 zi 2y 3 z 2 j + xy 2 zk) (1.17-a)
x y z
x2 z 2y 3 z 2 xy 2 z
= + + (1.17-b)
x y z
= 2xz 6y 2 z 2 + xy 2 (1.17-c)
= 2(1)(1) 6(1) (1) + (1)(1) = 3 at (1, 1, 1)
2 2 2
(1.17-d)
A.1.3 Gradient
A.1.4 Scalar
17The gradient of a scalar eld g(x) is a vector eld g(x) such that for any unit vector v,
the directional derivative dg/ds in the direction of v is given by
dg
= gv (1.20)
ds
where v = dp
VECTOR OPERATIONS
ds We note that the denition made no reference to any coordinate system. The
gradient is thus a vector invariant.
18 To nd the components in any rectangular Cartesian coordinate system we use
dp dxi
v = = ei (1.21-a)
ds ds
dg g dxi
= (1.21-b)
ds xi ds
which can be substituted and will yield
g
g = ei (1.22)
xi
or
i+ j+ k (1.23-a)
x y z
= i+ j+ k (1.23-b)
x y z
and note that it denes a vector eld.
19 The physical signicance of the gradient of a scalar eld is that it points in the direction in
which the eld is changing most rapidly (for a three dimensional surface, the gradient is pointing
along the normal to the plane tangent to the surface). The length of the vector ||g(x)|| is
perpendicular to the contour lines.
20 g(x)n gives the rate of change of the scalar eld in the direction of n.
Determine the gradient of = x2 yz+4xz 2 at point (1, 2, 1) along the direction 2ij2k.
Solution:
The stress tensor throughout a continuum is given with respect to Cartesian axes as
A7
3x1 x2 5x22 0
= 5x22 0 2x23 (1.25)
0 2x3 0
Determine the stress vector (or traction) at the point P (2, 1, 3) of the plane that is tangent
to the cylindrical surface x22 + x23 = 4 at P , Fig. A.6.
x3
n
x2
2 3
1
x
1
Solution:
v
vx y vz
x x x
vx vy vz
[x v] = y y y (1.31)
vx vy vz
z z z
vx vx vx
x y z
vy vy vy
[vx ] = x y z (1.32)
vz vz vz
x y z
that is [v]ij gives the rate of change of the ith component of v with respect to the jth
coordinate axis.
22 Note the diference between vx and x v. In matrix representation, one is the transpose
of the other.
23 The gradient of a vector is a tensor of order 2.
24 We can interpret the gradient of a vector geometrically, Fig. A.7. If we consider two points
a and b that are near to each other (i.e s is very small), and let the unit vector m points in
the direction from a to b. The value of the vector eld at a is v(x) and the value of the vector
eld at b is v(x + sm). Since the vector eld changes with position in the domain, those two
vectors are dierent both in length and orientation. If we now transport a copy of v(x) and
place it at b, then we compare the dierences between those two vectors. The vector connecting
the heads of v(x) and v(x + sm) is v(x + sm) v(x), the change in vector. Thus, if we
divide this change by s, then we get the rate of change as we move in the specied direction.
Finally, taking the limit as s goes to zero, we obtain
v(x+ s m ) -v(x)
v(x+ s m )
v(x)
x3
a sm b
x2
x1
The quantity Dv(x)m is called the directional derivative because it gives the rate of
change of the vector eld as we move in the direction m.
d
if a vector S(u) exists such that R(u) = du (S(u)), then
d
R(u)du = (S(u)) du = S(u) + c (1.35)
du
26 Given r(u) = x(u)e1 + y(u)e2 + z(u)e3 where r(u) is a position vector dening a curve
C connecting point P1 to P2 where u = u1 and u = u2 respectively, anf given A(x, y, z) =
A1 e1 + A2 e2 + A3 e3 being a vectorial function dened and continuous along C, then the integral
of the tangential component of A along C from P1 to P2 is given by
P2
Adr = Adr = A1 dx + A2 dy + A3 dz (1.36)
P1 C C
If A were a force, then this integral would represent the corresponding work.
27 If the contour is closed, then we dene the contour integral as
,
Adr = A1 dx + A2 dy + A3 dz (1.37)
C C
31 The divergence theorem (also known as Ostrogradskis Theorem) comes repeatedly in solid
That is the vector divergence over a volume is equal to the vector ux over a surface.
32 For 2D-1D transformations, we have
,
qdA = qT nds (1.42)
A s
If we select vT = [ 0 0 ], we obtain
d = nxd d (1.44)
x x
Solution:
A uid has a velocity eld v(x, y, z) and we rst seek to determine the net inow per unit
time per unit volume in a parallelepiped centered at P (x, y, z) with dimensions x, y, z,
Fig. A.8-a.
Z
D
E
Z V Y
C
V A P(X,Y,Z) H F V
V
X
B
Y G
a)
S
X
n dV=dxdydz
dS
Vt n
dS
b)
c)
vx |x,y,z vx (1.47-a)
1 vx
vx xx/2,y,z vx
x AFED (1.47-b)
2 x
1 vx
vx x+x/2,y,z vx + x GHCB (1.47-c)
2 x
The net inow per unit time across the x planes is
1 vx 1 vx
Vx = vx + x yz vx x yz (1.48-a)
2 x 2 x
vx
= xyz (1.48-b)
x
Similarly
vy
Vy = xyz (1.49-a)
y
vz
Vz = xyz (1.49-b)
z
Hence, the total increase per unit volume and unit time will be given by
' (
vx vy vz
x + y + z xyz
= div v = v (1.50)
xyz
Victor Saouma Finite Elements II; Solid Mechanics
Draft
A12 VECTOR OPERATIONS
Furthermore, if we consider the total of uid crossing dS during t, Fig. A.8-b, it will be given
by (vt)ndS = vndSt or the volume of uid crossing dS per unit time is vndS.
Thus for an arbitrary
volume, Fig. A.8-c, the total
amount of uid crossing a closed surface
S per unit time is vndS. But this is equal to vdV (Eq. 1.50), thus
S V
vndS = vdV (1.51)
S V
Appendix B
CASE-STUDY: FRACTURING of
A DAM DUE TO THERMAL
LOAD
B.1 INTRODUCTION
The dam is a 60 m high arch gravity dam built in the early fties. In its original conguration,
the crest supported a 5 m wide road bridge. There is no indication of any malfunction until the
ood of 1981 occurred. At that time, the bridge abutment were clogered by trees and debris,
preventing the eective overtopping of the water.
Following that incident, it was decided to dismantle the bridge from the dam crest, and a
new one, parallel to the crest chord, was built.
Soon after, a crack was observed inside the entire upper gallery, on the downstream face.
Crack mouth opening displacement were recorded, and crack maps drawn. Unfortunately,
readings were not taken systematically at the same pool elevation (which varies by +/- 10 m),
and no clear trend can be observed. However, it should be noted that the crack opening is
stable and there is no indication of increased values. Finally, no observation were conducted on
the downstream face to detect the presence of any daylighting crack.
Time day
h
CASE-STUDY: FRACTURING of A DAM DUE TO THERMAL LOAD
1 m Thickness
2,400 Kg/m3 Mass density
E 36 109 Pa Youngs modulus
0.2 Poissons ratio
1 105 m/m/o C Coecient of thermal expansion
hair 34 W/m2 o C Coecient for heat transfer by convection, air
hwat 100 W/m2 o C Coecient for heat transfer by convection, water
briey described in the Appendix, and the selected properties are shown in Table B.3.
B.1.2 Loads
Loads adopted in the analysis are shown in Table B.4.
of hydration caused by the 280 Kg/m3 concrete caused a thermal gradient large enough in the
downstream face to induce cracking.
Hence, our rst step was to create a boundary le which accounts for the dierent concrete
mixes used in in the dam, Fig. B.1
2. Assume that at the interfaces 6-3 and 8-5 the temperature is 25o C, whereas 7-4 it is 20o C
3. Assume that the outside temperature, and the one of the gallery is equal to 0o C
5. Perform a transient analysis with time increment equal to one day. Heat of hydration
values are taken from day 10 to 60.
35 34 33 56
59 58 57 32
31
36 55
60 54
37 9
61 51 18 52 2953 30
19 49
30 24
34
35 50 48 8
6 31 29 7 33
36
15 28 46 32 47 17
18 14 22 16 26 25 40 41
3 39
17 23 25 24 42
4 27 26
19 21 5
38 27 23 43
16 20 24 28 22
13 10
14 12 13 11 12
37 44
15 2 11
10 20 45 21
8 7
7
8
6
9 1
9
6
1 2
1
2 5
2 3 3 4 4 5
Hb = Hc mz /b
Age Hc Hc [J/Kg.day]
[days] [J/g] [J/g.day] mz
180 280
0 0
1.5 85.000 6,375 9,917
3 255
5 20.000 1,500 2,333
7 335
17.5 3.09524 232.14 361.11
28 400
59 0.48387 36.29 56.45
90 430
227.5 0.10909 8.18 12.73
365 460
1,368.5 0.01495 1.12 1.74
2,372 490
Table B.6: Heat of Hydration Adopted in the Simulation; Days and J/Kg/Day
Contour Plot,
Temperatures, Temperatures
3.00e+1
2.70e+1
2.40e+1
2.10e+1
1.80e+1
1.50e+1
1.20e+1
9.00e+0
6.00e+0
3.00e+0
Y
X -2.25e-3
Z
When such an analysis was performed, it was observed that the concrete temperatures were
very small (about 1-2o C). This was attributed to the lack of incremental analysis, of having one
single large lift being investigated, and with an initial temperature of zero, not enough heat
was released to warm the temperature. Clearly this is erroneous.
To qualitatively alleviate for this limitation (caused by lack of incremental analysis), and
keeping in mind that we are primarily interested at this point in determining qualitatively a
mechanism which may cause cracking, the specic heat was reduced by a factor of 10 (no other
reduction factors were attempted).
With such a set of values, the largest temperature gradient was observed to occur at day 8
with the temperature distribution shown in Fig. B.3. We observe the thermal gradient along
the crack line.
Contour Plot,
Principal Stresses, Maximum
1.19e+7
1.05e+7
9.03e+6
7.58e+6
6.14e+6
4.69e+6
3.24e+6
1.80e+6
3.53e+5
-1.09e+6
Y
X -2.54e+6
Z
Crack Lig. n t
0.000E+00 -1.136E+06 -1.341E+06
4.168E-01 -8.677E+05 -1.264E+06
6.947E-01 -9.974E+05 -9.699E+05
1.228E+00 -1.112E+06 -7.180E+05
1.317E+00 -1.123E+06 -7.853E+05
1.690E+00 -7.825E+05 -1.284E+06
1.873E+00 1.907E+05 -6.107E+05
2.205E+00 5.690E+05 -5.777E+05
2.537E+00 8.226E+05 -5.414E+05
2.664E+00 7.725E+05 -4.123E+05
2.996E+00 7.024E+05 -3.491E+05
3.059E+00 6.063E+05 -2.992E+05
3.455E+00 5.663E+05 -2.918E+05
3.455E+00 4.615E+05 -2.169E+05
3.959E+00 3.259E+05 -1.364E+05
4.296E+00 1.699E+05 -6.244E+04
4.633E+00 3.613E+04 -1.220E+04
4.969E+00 3.750E-04 -1.098E-04
B.3 CONCLUSIONS
The following conclusions may be drawn from this preliminary analysis
1. There are two potential reasons for which cracking might have occurred:
(a) Thermal gradient between cold water and warm downstream air temperature ag-
gravated by the removal of the bridge which was acting as a corbel on a cathedral
buttress (i.e. its dead weight was osetting exural tensile stresses). The analysis
conducted to investigate this scenario was quite reliable quantitatively.
(b) Thermal shock during construction. The analysis conducted toward such a cause is
mostly qualitative. A more rigorous incremental transient thermal analysis would be
needed to further support this theory. However, this analysis may be quite complex
and may be outside the scope of this investigation.
2. Dierential settlement was ruled out as there was no indication of foundation measurable
excessive deformation.
3. A three dimensional analysis has been initiated (hydrostatic load), but due to limitation
of the authors PC, it was not extended to the thermal analysis. It would be interesting
to contrast results of such an analysis with its 2D counterpart.
5. It would be interesting to determine if indeed upstream cracks are present in the eld (as
this investigation predicts).
Appendix C
MISC.