See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/258217834

Selective mass scaling and critical time-step estimate for explicit dynamic
analyses with solid-shell elements

Article · October 2013

DOI: 10.1016/j.compstruc.2012.10.021

CITATIONS READS

43 1,002

3 authors:

Giuseppe Cocchetti Mara Pagani

Politecnico di Milano COMSOL Srl
63 PUBLICATIONS   649 CITATIONS    23 PUBLICATIONS   112 CITATIONS   

SEE PROFILE SEE PROFILE

Umberto Perego
Politecnico di Milano
172 PUBLICATIONS   1,559 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Blade cutting of thin-walled structures View project

numerical simulation of landslide runout View project

All content following this page was uploaded by Umberto Perego on 14 March 2019.

The user has requested enhancement of the downloaded file.

 Computers & Structures, Vol. 127 (October 2013), pp. 39-52, doi:10.1016/j.compstruc.2012.10.021
Selective mass scaling and critical time-step estimate for explicit dynamics
analyses with solid-shell elements

Giuseppe Cocchettia , Mara Pagania , Umberto Peregoa,∗

a
Department of Structural Engineering, Politecnico di Milano, Piazza L. da Vinci,32, 20133 Milan, Italy

Abstract
Explicit integration is often used in highly nonlinear finite element structural dynamics simulations.
However, explicit time integration is stable only if the used time-step is smaller than a critical threshold,
which can be shown to depend on the smallest geometrical dimension of the finite elements in the mesh.
This aspect is particularly critical when solid-shell elements are used for the analysis of thin walled
structures, since the small thickness can lead to unacceptably small time-steps. A selective mass scaling
technique, based on a linear transformation of the element degrees of freedom, is proposed in this paper
to increase the size of the critical time-step without affecting the dynamical response. An analytical
procedure is also developed for the computation of the element highest eigenfrequency and estimate of
the critical time-step size. The computational effectiveness and accuracy of the proposed methodology
is tested on the basis of numerical examples.
Keywords: Explicit Time Integration; Selective Mass Scaling; Solid-Shells.

1. Introduction

Thin walled structures appear in many important engineering applications. For effective and accurate
computations, structures of this type are mostly analyzed using shell finite elements, belonging to either
one of two main categories: shell elements derived on the basis of the classical or degenerate shell
concepts (see[1]), in conjunction with the assumption of plane stress state (see e.g. [2–4]); solid-shell
elements, directly derived from three-dimensional continuum elements, not using rotational degrees of
freedom and allowing for the implementation of fully three-dimensional constitutive laws (see, e.g., [5, 6]
and [7], part I and part II).
In recent times, solid-shell elements have been studied more and more intensively, since they are
claimed to present several advantages over classical shell elements. Common arguments used in the
literature in favor of this statement are: more straightforward enforcement of boundary conditions, pos-
sibility to incorporate complex 3D material models, no need for complex update algorithms for finite
rotations (event though recent advancements, such as the rotation vector parametrization proposed e.g.
in [8] have greatly reduced the complexity of the problem), easy usage in combination with 3D solid el-
ements, possibility to obtain good accuracy in the through-the-thickness stress distribution in laminated
composites, thanks to direct modeling of thickness strains. On the other hand, low-order continuum
elements, from which solid-shell elements are usually derived, exhibit several types of locking behavior
which, however, can now be controlled adopting assumed strain and/or enhanced strain corrections of
the element kinematics. This aspect will not be considered in this paper and the reader is referred to the
recent literature on the subject (see, e.g., [9–12] and references therein ).

∗
Corresponding author
Email addresses: giuseppe.cocchetti@polimi.it (Giuseppe Cocchetti), pagani@stru.polimi.it (Mara Pagani),
umberto.perego@polimi.it (Umberto Perego)

Preprint submitted to Elsevier January 14, 2015

 Despite these recent achievements in solid-shell formulations, the fact that the thickness dimension is
always significantly smaller than the in-plane dimensions unavoidably leads to a high ratio of transverse
to in-plane normal stiffnesses, with a high finite element maximum eigenfrequency and, hence, stiff-
ness matrix ill-conditioning. These inherent difficulties are particularly relevant when iterative solvers
are used in implicit formulations or when explicit time integration is employed in dynamic analyses.
Explicit integration is usually adopted in wave propagation problems. However, even though implicit in-
tegration is the most natural approach in the case of inertia dominated problems, because of convergence
difficulties which may occur in highly nonlinear problems (such as, e.g., in contact problems), explicit
integration is often preferred also in this case.
The present contribution is therefore focused on explicit central difference approaches, which are
only conditionally stable. In the undamped case, a stable time-step must satisfy the condition
2
∆t ≤ (1)
ωmax
where ωmax is the maximum eigenfrequency of the assembled mesh. One can also show that

ωmax ≤ ωemax (2)

where ωemax is the highest eigenfrequency of an individual element in the mesh. A physical interpretation
of the stability limit can be obtained from the Courant, Friedrichs and Lewy condition which prescribes
that the time-step must be smaller than the time required by a dilatational stress wave to traverse the
smallest element of the mesh s
Le λ + 2µ
∆t ≤ , c = (3)
c ρ
where Le is the element smallest geometric dimension (element characteristic length), λ and µ are Lame’s
constants, ρ is the mass density and c is the wave propagation speed. The ratio Le /c will be henceforth
referred to as “element traversal time”.
Two provisions have been recently proposed to circumvent the problem of solid-shell ill conditioning:
thickness scaling and mass scaling. In the first case [13–15], the stiffness matrix is modified so as to
improve its conditioning. In this case the scaling has either to be closely incorporated into the shell
formulation, or it is obtained through multiplication by a preconditioner, which is computationally not
effective in explicit analyses. In the second case, the mass matrix is modified, so as to reduce the speed
of dilatational stress waves traveling through the shell and hence increase the critical time-step in explicit
dynamics. In the present contribution a selective mass scaling technique for explicit dynamics analyses
employing solid-shell elements is proposed and assessed on the basis of numerical tests.
The simplest mass scaling technique consists of an artificial uniform increase of the material density,
which affects in the same way all structural eigenmodes, and therefore leads to unacceptable inaccuracies
in most cases. More accurate approaches can be pursued considering that individual finite elements con-
tribute to the lowest structural eigenmodes mainly with the inertia associated to their rigid body modes.
In inertia dominated problems, better results can then be obtained by selectively scaling element masses,
in such a way that masses associated to element rigid body modes are not modified. A theoretically moti-
vated scaling, which satisfies this requisite, can be obtained by summing to the mass matrix the stiffness
matrix multiplied by a scaling parameter [16, 17]. This scaling can be shown to selectively reduce the
higher structural eigenfrequencies, with little or zero modifications of the lowest ones. The price to pay
is that, after the scaling, the originally lumped mass matrix becomes non-diagonal. The discretization
of complex structures generally leads to non-uniform meshes, with fine meshes often localized in small
regions, which govern the time-step size for the whole system. In this cases, the scaling can be applied
to relatively small patches of elements and the non-diagonal mass matrix can be effectively solved for
the nodal accelerations using an iterative scheme [18]. In the general case of thin walled structures,
however, the uniform small thickness is the dominant geometric factor and the scaling should be applied
to the whole mesh, leading to excessive computational costs when the mass matrix has to be inverted in
2
explicit approaches. Other mass scaling techniques are discussed in Hetherington and Askes [19] and
Askes et al. [20], all leading to non-diagonal mass matrices.
To avoid loosing the mass matrix diagonal structure, degrees of freedom governing the element rigid
body modes should be segregated and the scaling applied to the remaining degrees of freedom only. This
is the concept at the basis of the scaling technique usually adopted in classical shell elements [21]. Since
the element rigid body modes are governed by the middle plane degrees of freedom, the idea is to increase
the masses associated to rotational degrees of freedom only. In solid-shell elements, only displacement
degrees of freedom are used, and the technique cannot be used in a straightforward way. The technique
discussed in this paper is then based on a linear transformation of the degrees of freedom, which allows
to selectively apply the mass scaling while preserving the mass lumping. It is shown how this can be
accomplished in a simple and computationally inexpensive way, so that the time-step size turns out to be
governed by the element in-plane dimensions only (element in-plane traversal time), independent of the
element thickness.
After the mass scaling is carried out, one needs to accurately compute (or to bound from below)
the critical time-step. This can be obtained by estimating the maximum eigenfrequency of each finite
element in the mesh. Gershgorin’s theorem [22] is widely used to compute an upper bound to such an
eigenfrequency. However, when selective mass scaling is applied, the bound turns out to be overly large,
leading to too small time-steps. For eight-node hexaedron with constant strains (one integration point)
and uniform density, Flanagan and Belytschko [23] provided closed form expressions for bounding the
maximum eigenfrequency both in undistorted and distorted element shapes. The critical time-step can
also be obtained as the time required by a dilatational stress wave to traverse the smallest element di-
mension which, for solid-shell elements, is always the thickness. When the element is distorted, the
characteristic length can be estimated from below on the basis of geometric considerations [24], but this
technique can hardly be used in a consistent way in conjunction with selective mass scaling. Alterna-
tively, ωemax can be found solving analytically the eigenvalue problem. Besides the solutions provided
in Flanagan and Belytschko [23], Ling and Cherukuri [25] provided analytical solutions for plane stress
and strain four and eight-node quadrilaterals. Comini and Manzan [26] provided analytical solutions,
for conduction type problems, for various two and three-dimensional elements, including triangles and
tetrahedra.
In the present work, it is shown how the eigenvalue problem of the undistorted parallelepiped with
selectively scaled masses can be reduced to a sequence of second or third order polynomial equations.
Explicit expressions for ωemax are provided and several numerical examples are used to test the accuracy
and computational effectiveness of the proposed selective scaling technique. It is also shown how the
analytical computation of ωemax can lead to significant computational gains in some circumstances.

2. Selective mass scaling

The dynamic equilibrium equations of the undamped discretized system are given by

Ma + f int = f ext (4)

where a is the vector of nodal accelerations, M is the mass matrix, f int and f ext the vectors of equivalent
internal and external nodal forces, respectively. The effect of prescribed displacements is assumed to be
incorporated in f ext . The implementation of the central difference integration scheme requires that the
accelerations are computed at each time-step as

a = M−1 f (5)

where f = f ext − f int and, for effective computations, M is assumed to be diagonal.

As anticipated in the previous section, the idea is to modify the element mass matrix preserving its
diagonal structure and without affecting lower order structural eigenmodes. This can be achieved by
scaling masses of individual elements, in such a way that the energy associated to their rigid body modes
remains unaltered. In solid-shell elements the thickness is inherently different from the element in-plane
3
 Figure 1: Solid-shell element: definition of upper and lower surfaces.

dimensions and can be easily identified from the shell configuration. Making reference to the eight-node
solid-shell element with lumped masses shown in Figure 1, the upper and lower surfaces of the element
up
i , i = 1, ...4, denote the accelerations of corresponding pairs of
can also be easily identified. If ai and alow
nodes on the two surfaces, the accelerations aavei governing the element translational rigid body modes
can be defined using the following linear transformation
up
ai + alow
aave
i = i
(6)
2
diff
while the accelerations ai , governing the higher order modes, are defined as
up
diff a − alow
ai = i i
(7)
2
with the inverse transformations
up diff diff
ai = aave
i + ai , alow
i = aave
i − ai (8)

For element e, equations (8) can be cast in matrix form as

 up e #e  e

 a   "
I I

 aave 

a =
e
= Tâ =
e
 12×1
 
 12×1 
(9)
  
low  I −I  adiff
 a12×1 
 
24×1 
 
 
 12×1 



where I is the 12 × 12 identity matrix. Introducing the transformation (9) in the virtual balance equation,
one obtains the expressions of the mass matrix and vector of equivalent nodal forces expressed in terms
of average and difference degrees of freedom (a superposed hat denotes transformed quantities):

 . . . ..
 e
 0 . 0

up up
 (mi + mlow )
i (mi − mlow
i ) 
.. . . 
 
)e
. .  f ave
 (
 0 0
M̂ = T M T = 
e T e  , f̂ = T f =
e T e
 . . . .. f diff
.

0 0 
up up
(mi − mlow (mi + mlow

) i )

i
 
 .. . . 
0 . 0 .
(10)
up
where mi and mlow i are the 3 × 3 matrices containing the nodal masses at the i − th pair of nodes
up
(i = 1, . . . 4). For distorted elements or non-uniform density (i.e. when mi , mlow e
i ), M̂ is not diagonal
e
even if the original matrix M is lumped. However, as it appears from equation (10), the non-diagonal
4
terms are expected to be small and a diagonal mass matrix M̂elumped can be obtained by simply neglecting
them. Note that using, e.g., the HRZ lumping procedure [27] would lead to the same result. On the
other hand, the transformed mass matrix is rigorously diagonal for elements of parallelepiped shape with
uniform density. In view of these considerations, a lumped transformed mass matrix M̂elumped will be
used in what follows:
 .
 . .

 ave e 0 
 M 0   up

M̂elumped =  12×12 diff
 , Mave = Mdiff =  (mi + mlow )
i
 (11)
0 M   . . 

12×12
0 .

At this point, the element maximum eigenfrequency can be scaled down, without affecting the element
rigid body modes, leaving Mave unaltered and multiplying Mdiff by a scale factor α > 1 defined as
#e
Aup + Alow
" ave
ˆ M 0
α=β , M̄lumped =
e
(12)
2h2 0 αMdiff

where Aup and Alow are the areas of the element upper and lower surfaces and h = 41 4i=1 hi is the average
P
element thickness, hi being the distance between corresponding nodes on the upper and lower surfaces
(Figure 1). The factor β can be adjusted in such a way that the scaled critical time-step size approaches
the in-plane element traversal time.
To obtain a reduction of the maximum eigenfrequency, the new element degrees of freedom with
scaled masses need to be used also for the global assembled mesh. This is not a problem since they are
nodal degrees of freedom and can be assembled with the usual procedure. If the transformed element
matrices are assembled in the equation of motion, one simply obtains
ˆ āˆ = f̂
M̄ (13)

where a superposed bar denotes the scaled mass matrix and relative nodal accelerations. Referring to the
diff
j − th node in the mesh, once āave
j and ā j have been computed as

up diff up
f ave
j f j + f low
j diff
fj f j − f low
j
āave
j = up = up , ā j = up = up (14)
m j + mlow
j m j + mlow
j α(m j + mlow
j ) α(m j + mlow
j )

one can easily recover the nodal accelerations in terms of the original degrees of freedom
up
up diff
(α + 1)f j + (α − 1)f low
j
ā j = āave
j + ā j = up
α(m j + mlow
j )
up (15)
diff
(α − 1)f j + (α + 1)f low
j
ālow
j = āave
j − ā j = up
α(m j + mlow
j )

Remark 1. A conceptually almost identical scaling procedure was presented in Olovsson et al. [28]. In
that case, rather than to the masses, the scaling was applied directly to nodal accelerations. Considering
diff
a node on the upper surface and using the present notation, Olovsson et al. [28] defined a j as the
up up
difference between the original, unscaled acceleration f j /m j and the rigid body acceleration aavej , i.e.

up up up up up
diff
fj fj f j + f low
j mlow low
j fj − mj fj
aj = up − aave
j = up − up = up up (16)
mj mj m j + mlow
j m j (m j + mlow
j )

diff
and then applied the scaling to a j directly, obtaining the scaled nodal acceleration

up 1 diff
ā j = aave
j + a , α≥1 (17)
α j
5
 Figure 2: Reference element for analytical computation of maximum eigenfrequency.

If the same procedure is repeated for the corresponding node on the lower surface, a different expression
diff up up diff
is obtained for a j , unless m j = mlow j . If m j = m j , a unique definition of a j is obtained also in
low

Olovsson et al. [28] case, coincident with the definition (7) given here. In this case the two procedures
are therefore identical. The advantage of the procedure proposed here is that it provides a consistent
variable transformation also for the stiffness matrix, so allowing for the analytical computation of ωemax
which will be discussed in the next section.
Remark 2. The formulation of a solid-shell element in terms of average and difference degrees of
freedom was originally proposed in Büchter and Ramm [29] and then considered in several other papers,
including Wall et al. [13], Gee et al. [14], Klöppel et al. [15], where procedures involving the scaling of
the vector of internal forces were discussed. In particular, in Wall et al. [13] an application to explicit
dynamics was considered.
Remark 3. Even though the scaling procedure discussed above is conceptually based on the degrees
of freedom transformation (9), its implementation does not require any modifications in the element
subroutines of an existing code, since mass scaling is applied directly to nodal accelerations, as shown
in (15). The only requirement is that care should be put in the identification of nodes belonging to the
upper and lower element surfaces. As it will be shown in the next section, also the procedure for the
computation of the time-step size is based only on the element geometry, and does not require to actually
implement the transformation (9).

3. Analytical estimate of element maximum eigenfrequency

3.1. Complete set of element eigenfrequencies

The implementation of the central difference integration scheme requires to compute a lower bound
to the critical time-step size. As shown in (1) and (2), this requires the computation of the element
maximum eigenfrequency ωemax . An upper bound to ωemax can be computed using Gershgorin’s theorem
as:
24
X |Ki j | 2
ωemax ≤ ωeGershgorin = max , ∆tGershgorin = e (18)
i
j=1
mi ωGershgorin

where Ki j are the entries of the i−th row of the element stiffness matrix and mi the corresponding lumped
mass. Unfortunately, as it will be shown in the examples, Gershgorin’s bound may become overly large
when selective mass scaling is applied, leading to excessive computational costs. This can be easily
explained, noting that, when mass scaling is applied to an element formulated in terms of average and
difference degrees of freedom, one has
 
24 24
| K̂ | | K̂ |

 

i j i j
X X
ω̄Gershgorin = max 
e
,
 
max max (19)
 
i ∈ ave j=1 m̂i i ∈ diff j=1 αm̂i 

 



i.e., for α sufficiently large, ω̄eGershgorin is always associated to a degree of freedom average, and does not
decrease for increasing scaling factor. This will be verified in the second example of the next section.
In the simple case of an isotropic linear elastic finite element with a regular parallelepiped shape, with
thickness h = 2c significantly smaller than the in-plane dimensions 2a and 2b (see Figure 2), the element
6
maximum scaled eigenfrequency ω̄max can be computed analytically following the procedure described
below. Note that to simplify the notation, the element index e will be omitted from now onward.
Considering a solid finite element with lumped and selectively scaled masses, unconstrained and
unloaded, in view of the system linearity, its equation of motion can be written as
ˆ
M̄ ˆ + K̂d̄ˆ = 0
lumped ā (20)

where K̂ is the stiffness matrix of a fully integrated (eight Gauss points), fully compatible 8-node brick
element, expressed in terms of average and difference degrees of freedom, āˆ and d̄ˆ are its nodal acceler-
ations and displacements and, again, a superposed bar denotes quantities affected by the mass scaling.
Assuming a solution in the form d̄ˆ = Ū ˆ sin (ω̄t), Ū
ˆ being an element eigenmode, one has that the
element eigenfrequencies are obtained from the eigenproblem
ˆ ˆ
 
lumped ω̄ − K̂ Ū = 0
2
M̄ (21)

Note that the linear transformation in (9) does not alter the element eigenfrequencies. This explains why
ω̄ in (21) does not have a superposed hat.
The solution of the eigenproblem in (21) is simplified if the following normalization of the mass and
stiffness matrices is adopted:

ˆ cE
M̄ lumped = ζ M M, K̂ = ζK K, ζ M = 2ρabc, ζK = (22)
72(1 + ν)
where E is Young’s modulus and ν is Poisson’s coefficient. Using these definitions, the eigenproblem is
redefined in nondimensional form as
ζM 2 (1 + ν) 2
(Mη − K)U = 0, η= ω̄ = 144 abρ ω̄ (23)
ζK E
The element aspect ratio will also enter the eigenproblem solution through the ratios
c c
γ= ; λ= (24)
a b
In practical terms, one is required to find the highest root of the 24-th order characteristic polynomial
of (23)1 . Since the element is unconstrained, the first three eigenmodes are rigid body translations and
the corresponding eigenvalues are equal to zero. The remaining 21 eigenvalues, including the three ones
corresponding to rigid body rotations, which are also equal to zero, are determined by considering sepa-
rately the eigenproblems corresponding to the groups of eigenmodes shown in Figure 3. The eigenmodes
have been grouped on the basis of their symmetry properties, which have been hinted at on the basis of
a complete numerical solution of the eigenproblem. The solution procedure for Group 3, which is the
one containing the eigenmode 9, corresponding to the highest eigenfrequency, is sketchily summarized
below and in Appendix 1. For the sake of completeness, the equations returning the eigenvalues of the
other groups are reported in Appendix 2.
Eigenmodes of Group 3 (See Figure 3b)
Due to symmetry requirements, the eigenvector U must take the following form:

U7,8,9 = {−u x u x u x − u x | −uy − uy uy uy | 0| 0{z0 }0 | 0| 0{z0 } 0 | 0| 0{z0 } 0 | uz uz uz uz }T (25)

| {z } | {z } ave
| {z }
ave
Ux Uyave Uz diff
Ux
diff
Uy diff
Uz

where all entries are expressed in terms of the three parameters u x , uy and uz only. The entries ordering
can be easily understood with the help of Figure 2. If the expression of U in (25) is substituted in the
eigenproblem (23)1 , u x , uy and uz can be solved in terms of the parameter η by taking any three equations,
two of which linearly independent, from the resulting linear homogeneous system, obtaining

7
Figure 3: Element eigenmode Groups. Dashed lines denote undeformed configuration. Solid lines denote eigendeformation.
Eigenmodes 16, 18 and 20 are rigid body rotations corresponding to a zero eigenvalue.

8
 u x = 144 ν (γ η − 144 λ)
uy = 144 ν (λ η − 144 γ)   (26)
uz = η2 γ λ (1 − 2ν) − 144η γ2 + λ2 (1 − ν) + 1442 γ λ
where η solves the following cubic equation:
h  i
α γ2 λ2 (1 − 2ν) η3 − 144 γ λ 1 + α γ2 + λ2 (1 − ν) η2 +
  (27)
+1442 λ2 + γ2 + α γ2 λ2 η − 1443 γ λ (1 + ν) = 0

The roots of (27) can be derived explicitly using a standard procedure. For convenience, since it will be
used in the next section, it has been summarized in Appendix 1.
Remark 4. For solid-shell elements, where the thickness dimension is significantly smaller than the
in-plane ones, the highest element eigenfrequency always turns out to be given by the square root of
the highest eigenvalue of Group 3, corresponding to thickness vibrations (eigenmode 9 in Figure 3b).
Numerical experiments, carried out for different combinations of c/a, c/b, ν and α, show that this is
always true, for all non-negative values of Poisson’s coefficient, for all ratios of c/a and c/b, as long as
c is the smallest dimension, and for all values of the mass scaling parameter α ≥ 1. Increasing the mass
scaling factor emphasizes the predominance of this eigenfrequency. However, so far we have not been
able to define analytical conditions for this to be true.
Remark 5. The element highest eigenvalue is the highest root of equation (27) which, for α = 1,
is identical to the cubic equation obtained by Flanagan and Belytschko in [23]. When selective mass
scaling is adopted (α > 1), the cubic equation modifies as in (27) and the estimates proposed in [23]
cannot be used anymore. This motivates the use of the analytical computation of the eigenfrequency
which will be discussed in the next subsection. The possibility to develop upper bounds of the type
proposed in [23], to be used in the presence of selective mass scaling, is currently under study.
Remark 6. Since the vibration mode corresponding to the highest eigenfrequency is known, it could be
considered computing it using Rayleigh quotient. However, the eigenvector shown in (25) depends on
the three parameters u x , uy and uz . Even though for thickness vibrations one knows that uz , 0 (recall
that uz in (25) denotes the difference between out of plane displacements of the element upper and lower
surfaces), due to nonzero Poisson coefficient there are also in plane vibrations governed by the parameters
u x and uy . The correct ratio between u x , uy and uz is part of the solution of the eigenvalue problem and
depends on the mass scaling parameter, on Poisson coefficient and on the element aspect ratio. Therefore,
the exact computation of Rayleigh quotient turns out to be computationally more expensive than the
eigenvalue direct computation through the root of the cubic equation (27). On the other hand, the use of
an approximate expression of the eigenvector in Rayleigh quotient would produce a lower bound of the
eigenvalue, which would not be suitable for the prediction of a stable time step.

3.2. Accuracy of the analytical estimate

The results presented in the previous section concern a reference brick element, which consists of an
8-node compatible finite element, fully integrated and of parallelepiped shape. This type of elements are
known to suffer several types of locking behavior, particularly in the thin limit, which are eliminated in
solid-shell elements by using enhanced strain and assumed strain approaches of various kinds. Further-
more, solid-shell elements for explicit dynamics applications are usually based on reduced integration.
For these reasons, it can be said in general terms that commonly used solid-shell elements are less stiff
than the reference brick element. In particular, the highest eigenmode in most solid-shell elements con-
sists of a uniform thickness extension, which is exactly represented in the reference brick element. As
far as the lower eigenmodes are concerned, those of the reference brick element represent a conservative
upper bound. In conclusion, in most cases the highest eigenfrequency computed from (27) can be con-
veniently used in (1) for the definition of the time-step size. This is clearly not true when the thickness
behavior is enriched with additional degrees of freedom if they have a specific inertia attached to them
(see, e.g., the 7-parameter shell and solid-shell elements [30–32]), because in these cases higher order
modes would appear which cannot be correctly represented by the reference brick element.
9
 As an example, the numerically computed eigenfrequencies of the solid-shell element Q1STs of
Schwarze and Reese [10, 12] are compared in Figure 4 to the analytical eigenfrequencies of the fully
integrated 8-node brick of same size and material, obtained as the highest root of the cubic equation in
(27). The Q1STs element is underintegrated, with only one in-plane Gauss point and several Gauss points
in thickness direction. Furthermore, it is endowed with hourglass control and shear, Poisson thickness,
curvature and volumetric locking controls by a combination of assumed natural strain and enhanced
strain techniques. As it can be noted, the highest eigenfrequency is identical in the two cases. The
other eigenfrequencies of the Q1STs are always not greater than the corresponding value of the 8-node
brick. For the five modes involving shear locking, the shear locking prevention technology implemented
in the Q1STs leads to eigenvalues which are almost two orders of magnitude smaller. The highest
eigenfrequency for different element aspect ratios is also reported in Table 1. Again, the analytical value
of the highest eigenfrequency, computed for the 8-node brick, is always identical to the corresponding
numerical value of the Q1STs so that only one value is reported. It can also be observed that, as expected,
the highest eigenfrequency is almost insensitive to the element in-plane aspect ratio and that it scales
almost linearly with the element thickness.
Table 2 shows the corresponding results for different amounts of selective mass scaling. Again, in-
dependent of the amount of scaling, analytical and numerical highest eigenfrequencies turn out to be
identical. However, the highest eigenfrequency does not scale anymore linearly with the thickness. In
Table 2, the scaling parameter α has been chosen such as to compensate for the reduction of thickness,
so that the eigenfrequency remains the same even for a difference in the aspect ratio of one order of mag-
nitude. As already discussed in Remark 5, this makes the standard techniques for estimating the stable
time-step not applicable in the presence of selectively scaled masses and motivates the development of
the analytical estimate pursued in this paper.

Figure 4: Parallelepiped element of size 2a = 2b = 5 mm, h = 0.02 mm, with E = 1768 MPa, ν = 0.3, ρ = 3000 kg/m3 .
Unscaled masses. Squared dots: analytical eigenfrequencies of 8-node brick. Rhombic dots: numerical eigenfrequencies of
Q1STs element with two Gauss points in thickness direction. Eigenfrequencies corresponding to rigid body modes are not
shown.

3.3. Operative procedure for the computation of element maximum eigenfrequency

In the case of distorted elements, the estimate for the parallelepiped element cannot be safely used.
This is true not only for meshes of initially distorted elements, but also of elements of initially regular size
which become distorted due to large deformations. In this case, since the time-step size decreases with
the smallest element geometric dimension, a conservative estimate can be obtained making reference
to a parallelepiped which is contained in the deformed elements. For slightly deformed elements, a
possible operative procedure is outlined in Box 1. The procedure can be applied both for the definition
10
Table 1: Parallelepiped element of size 2a = 5 mm, with E = 1768 MPa, ν = 0.3, ρ = 3000 kg/m3 . Unscaled masses. Identical
highest analytical eigenfrequencies of 8-node brick and/or numerical eigenfrequencies of Q1STs element with two Gauss points
in thickness direction, for varying element aspect ratio.
Highest eigenfrequency
α a/b a/c analytical/numerical
1 1 250 8.906952E+07
1 1 50 1,781516E+07
1 1 25 8,909548E+06
1 1 5 1,794992E+06
1 2 250 8,906992E+07
1 2 50 1,781713E+07
1 2 25 8,913501E+06
1 2 5 1,818283E+06
1 2.5 250 8,907021E+07
1 2.5 50 1,781861E+07
1 2.5 25 8,916485E+06
1 2.5 5 1,838798E+06
1 5 250 8,907267E+07
1 5 50 1,783100E+07
1 5 25 8,942009E+06
1 5 5 2,136975E+06

Table 2: Same as in Table 1, but with selectively scaled masses.

Highest eigenfrequency
α a/b a/c analytical/numerical
22710 2 250 8,139424E+05
908 2 50 8,139424E+05
227 2 25 8,139424E+05
14535 2.5 250 1,006166E+06
581 2.5 50 1,006166E+06
145 2.5 25 1,006166E+06
3619 5 250 1,986817E+06
1145 5 50 1,986817E+06
36 5 25 1,986817E+06

of the initial time-step and run-time, whenever a stability test states that element distortion is such that
the integration is not stable anymore with the current time-step. In this latter case, the new time-step is
multiplied by a safety factor equal to 0.8 to ensure stability in the presence of highly nonlinear structural
response.
Since an unstable integration leads to spurious generation of energy, the run-time time-step recal-
culation is based on an energy check. Let Wn+1 ext , W int and W kin be the external, internal and kinetic
n+1 n+1
energies at time tn+1 , respectively, defined as
1  
ext
Wn+1 =Wnext + ∆dT fnext + fn+1
ext
2
1  
int
Wn+1 =Wnint + ∆dT fnint + fn+1
int (28)
2
1
kin
Wn+1 = (vn+1 )T Mvn+1
2

11
where
∆d =∆tn+ 1 vn+ 1
2
 2  (29)
vn+1 =vn+ 1 + tn+1 − tn+ 1 an+1
2 2

∆tn+1/2 is the current time-step size and v is the vector of nodal velocities. Energy conservation requires
that  kin  n o
Wn+1 + Wn+1
int ext 
− Wn+1 ≤ ε max Wn+1
ext
, Wn+1
int
, Wn+1
kin
(30)

where ε is a tolerance which has been set to 1 · 10−2 (see [33], chapt. 6.2.3).
Whenever the energy stability check (30) is not satisfied, a new time-step size to be used in the next
step is computed according to the procedure illustrated in Box 1.

4. Numerical examples

4.1. Clamped cantilever plate

The slender plate shown in figure 5 is clamped on one side and subjected to a transversal tip load on
the other side. The tip load is applied with its final value at time t = 0 and then it is kept constant. The
plate material is assumed to be linear elastic with Young’s modulus E = 1768 MPa, ν = 0.3, ρ = 3000
kg/m3 . The plate is 100 mm long and 20 mm wide. Two thicknesses have been considered: h = 0.2 mm
and h = 0.02 mm. In the first case, a surface tip load of 1.56 · 10−3 MPa has been applied through the
thickness, while a reduced tip load of 1.56 · 10−4 MPa has been applied to the thinner plate, so as to limit
the tip displacement. The plate has been discretized using the mesh of 80 SHB8PS solid-shell elements
[11] shown in Figure 5. The elements are of regular shape with in-plane dimensions 2a = 2b = 5 mm.

Figure 5: Clamped cantilever plate: geometry and loading.

12
• Perform energy check (equation 30)
• IF energy check not satisfied THEN
 Compute dimensions of effective parallelepiped:
◦ compute element upper up low
n A and o lower A surfaces n ando volume V
◦ define: A max = max A , A , A = min Aup , Alow
up low min

V
2ceff = max
A n o
2amax = max length of edges of Amin
n o
2aeff = min length of edges of Amin
Amin
2beff = max
2a
 Compute critical time-step for effective parallelepiped of dimensions 2aeff , 2beff , 2ceff
 Compute coefficients of cubic equation (see Appendix 1 and equation 27)
ceff ceff
γ = eff ; λ = eff
a b
C0 = −1443 γλ (1 + ν) 
C1 = 1442 λ2 + γ2 + αγ2 λ2
h  i
C2 = −144γλ 1 + α γ2 + λ2 (1 − ν)
C3 = αγ2 λ2 (1 − 2ν)
!2
C1 1 C2
p= −
C3 3 C3
!3
C0 2 C2 1 C1C2
q= + −
C3 27 C3 3 C32
 Compute ∆t (see equation  27 and Appendix 1):

q
 r
p ϕ 1 C
2
ϕ = arccos − q    ; η9 = 2 − cos
 
−
 p 3 3 3 3 C3
2 − 3 
 s 
2  a b ρ + ν) 
eff eff (1
∆t = 0.8 eff = 0.8 24


η9 E

ω̄max
ELSE
• Continue analysis with same time-step

Box 1: Procedure for run-time time-step computation.

13
 Figure 7: Clamped cantilever plate. Plate tip deflection vs. time. a) thickness h = 0.2 mm. b) thickness h = 0.02 mm.

Figure 6: Analytical eigenfrequencies of individual finite element of size 2a = 2b = 5 mm, h = 0.02 mm, with E = 1768 MPa,
ν = 0.3, ρ = 3000 kg/m3 and with scaled (rhombic dots) and unscaled masses (squared dots). Scaling parameter α = 4.17 · 105 .
Zero eigenfrequencies corresponding to rigid body modes are not shown.

For the definition of the scale factor α, a factor β = 20/3 has been used in (12) for both thicknesses.
This value has been determined by inspection so as to scale the element maximum eigenvalue to a value
which approaches the maximum element eigenfrequency associated to in-plane deformation modes. This
allows to obtain a critical time-step of the same order of the in-plane traversal time. The effect of selec-
tive mass scaling on the eighteen non-zero element eigenfrequencies is illustrated in Figure 6 for the 0.2
mm thick plate. The plotted eigenfrequencies are computed analytically for the fully integrated 8-node
brick element (see discussion in section 3.2). Scaled eigenfrequencies are denoted by rhombic dots. It
can be noted that the smallness of the thickness, if compared to the in-plane element dimensions, leads to
a sudden jump of the eigenfrequencies in correspondence to the transition from in-plane to out-of-plane
deformation modes. The scaling has the effect to smooth the eigenfrequency profile, leading to a maxi-
mum eigenfrequency which approximates the maximum in-plane unscaled value. The eigenfrequencies
associated to out-of-plane eigenmodes are greatly reduced, while those corresponding to in-plane modes
are only slightly affected, due to non-zero Poisson coefficient.
The results of the analyses in terms of time histories of the tip deflection, for a duration of 0.15

14
 Table 3: Clamped cantilever plate: critical time-steps.
h 0.2 0.02
2a/h 25 250
α 1 1
∆tanalytical 1.30 · 10−7 1.30 · 10−8
Unscaled
∆tGershgorin 2.61 · 10−7 2.61 · 10−8
∆tOlovsson 2.25 · 10−7 2.25 · 10−8
α 4.17 · 103 4.17 · 105
∆tanalytical 4.63 · 10−6 4.69 · 10−6
Selectively Scaled
∆tGershgorin 1.47 · 10−6 4.63 · 10−7
∆tOlovsson 1.45 · 10−5 1.45 · 10−5

sec., are shown in Figures 7a and 7b, for h = 0.2 mm and h = 0.02 mm respectively. For h = 0.2
mm, the responses with scaled and unscaled masses are completely superposed. The thick curve shows
the response obtained with a uniform increase of the material density (“mass pumping”). As it can be
seen, for both thicknesses a uniform mass pumping heavily affects the structural lower modes, leading
to unacceptable results. Also for h = 0.02, despite the very large scale factor α = 4.17 · 105 adopted,
the scaled and unscaled tip deflections are very close. In this latter case, the used stable time-step is
more than two orders of magnitude larger than the time-step to be used in the unscaled case, leading to a
substantial computational saving.
The critical time-steps computed using different techniques are listed in Table 3, for the two consid-
ered thicknesses and with unscaled and selectively scaled masses. The second row reports the element
aspect ratio. The third and seventh rows the adopted scale factor α.
∆tanalytical is the time-step computed according to the procedure in Box 1. In the numerical simulation
it is computed at the beginning of the analysis and then kept constant, because the energy stability
criterion is always satisfied with the initial time-step. Therefore, the time-step sizes in Table 3 are not
affected by the 0.8 safety factor, which is applied only upon time-step recalculation. In large strain
problems, where elements undergo large distortions, the time-step size has to be recomputed on the basis
of the actual element configuration. For moderate distortions, as in the next example, this can be done
using the procedure presented in section 3.3 (see Box 1).
∆tGershgorin is the time-step computed using Gershgorin’s upper bound on the maximum element
eigenfrequency; ∆tOlovsson is the time-step computed in terms of the Courant, Friedrichs and Lewy con-
dition on the element traversal time, applying the scaling factor to the element characteristic length as
proposed in Olovsson et al. [28]: √ e
αL
∆tOlovsson = (31)
c
It should be noted that while ∆tanalytical is computed on the basis of the reference element (fully com-
patible, fully integrated) discussed in section 3, ∆tGershgorin is computed using the stiffness matrix of the
actual element employed in the simulations, the 8-node solid-shell SHB8PS proposed in Abed-Meraim
and Combescure [11]. Since the reference brick element does not implement any provisions to avoid
locking, its response is generally stiffer than the SHB8PS, as discussed in section 3.2, while its maxi-
mum eigenfrequency is computed exactly. ∆tanalytical is therefore a very accurate conservative estimate
of the critical time-step. On the other hand, ∆tGershgorin is also a rigorous lower bound of the critical
time-step, which is reasonably accurate for the unscaled element, but which turns out to be overly con-
servative when masses are scaled, as discussed in section 3.1, and of little use in practical computations.
The analysis has also been run using progressively larger time-steps, until an unstable time integration
has been obtained. It has been found that the larger stable time-step for both h = 0.2 and h = 0.02 mm
is 1.33 · ∆tanalytical , which confirms that the analytical estimate is conservative, with a safety margin that
in this case is of about 33%. It should be noted that the time-step estimate in (31) leads in this case to an
unstable integration.
15
 Figure 8: Cylindrical shell hinged on two sides: finite element mesh, boundary conditions and load.

4.2. Cylindrical shell hinged on two sides

The cylindrical shell hinged on two sides and free on the other two sides, subjected to a transversal
concentrated center load, has been analyzed using the 16 × 16 SHB8PS element mesh shown in Figure 8.
The problem has been used in Wall et al. [13] to test the application of the scaled director conditioning
technique [14] in explicit dynamics simulations. The shell has a radius of 5 meters, a thickness of 0.01
meters and a center opening angle of 80◦ . The material is assumed to be linear elastic with a Young’s
modulus of 2.0 · 109 N/m2 , Poisson’s ratio 0.0 and mass density 1.0 · 105 kg/m3 . The load is applied with
a linear ramp from a zero initial value to 1 MN after 0.2 seconds, and then kept constant.
The analysis has been carried out using three different scaling factors α: α = 1.27 · 102 , α = 1.27 · 103
and α = 1.27·104 (see equation 12). A reference analysis has also been carried out with the finite element
code Abaqus, using a 32 × 32 mesh of fully integrated shell elements (S4 element type from Abaqus
element library). The results in terms of center-point displacement evolution are shown in Figure 9.
The center-point displacement evolution for the unscaled and the α = 1.27 · 102 analyses is very
similar for the whole analysis duration. With α = 1.27 · 103 there is good agreement up to the peak
displacement, also in the post-buckling regime. After the peak, the analysis with α = 1.27 · 103 exhibits
a displacement reduction, which is not observed in the previous curves. It should be noted however
that the same displacement reduction is exhibited by the Abaqus reference curve, which remains very
close to the α = 1.27 · 103 curve throughout the analysis. The last curve, with α = 1.27 · 104 , diverges
significantly from the others, meaning that the adopted mass scaling affects in an unacceptable way
the shell dynamic response. Snapshots of the shell deformation are shown in Figure 10 for the case
α = 1.27 · 103 . Another beneficial effect of the mass scaling can be further observed noting that the
α = 1.27 · 103 curve is smoother than the unscaled curve, meaning that the spurious higher frequencies
are reduced as a consequence of the adopted mass scaling.
The initial time step-size used with the different analyses is reported in Table 4. While in the unscaled
case the time-step is computed using Gershgorin upper bound, in the scaled cases it is computed using
the procedure explained in section 3.2. As in the previous example, also in this case Gershgorin bound
is not accurate enough when mass scaling is applied, returning the same value ∆tGershgorin = 4.64 · 10−4
for all the adopted values of α. From Table 4, it can be observed that the α = 1.27 · 103 mass scaling
provides a gain of almost two orders of magnitude with respect to the unscaled analysis. The Abaqus
time-step is also significantly smaller, but one should take into account that the Abaqus mesh is made of
elements that are two times smaller than those used in the α = 1.27 · 103 analysis. On the other hand, the
additional gain obtained with α = 1.27 · 104 is relatively small, and does not compensate for the accuracy
16
 Table 4: Cylindrical shell hinged on two sides: initial time-step size.
Unscaled α = 1.27 · 102 α = 1.27 · 103 α = 1.27 · 104 Abaqus-S4
∆t [sec] 6.71 · 10−5 7.96 · 10−4 2.52 · 10−3 3.09 · 10−3 ≈ 1. · 10−5

Figure 9: Cylindrical shell hinged on two sides: time evolution of center-point deflection with scaled and unscaled masses, for
different scaling factors.

loss.
The time-steps sizes shown in Table 4 are just initial values. Since the problem entails large defor-
mations, the time-step need be recomputed using the procedure outlined in Box 1, whenever the energy
stability criterion is violated. Despite the large shell deformation, the elements become only slightly
distorted, so that time-step recalculation is activated only once for the unscaled, α = 1.27 · 102 and
α = 1.27 · 103 cases, while it is never activated for the α = 1.27 · 104 case. In all cases the time-step
reduction upon recalculation is of the order of 25-30%, including the adopted safety factor of 0.8. The
energy evolutions in the α = 1.27 · 103 case are shown in Figure 11. From Figure 11b, it can be observed
that the energy error never trespasses the threshold, even though slightly before 0.3 sec it undergoes a
sudden jump. At this point, the time-step is recalculated and stability of the integration remains under
control, despite visible oscillations in the energy error.

5. Conclusions

Solid-shell elements are solid elements with translational degrees of freedom only, characterized by a
small thickness compared to the in-plane dimensions. Even though solid-shell elements are becoming in-
creasingly popular for implicit nonlinear simulations, their usage in explicit dynamics is computationally
ineffective due to the very small stable time-step resulting from the small thickness.
In the present paper, a selective mass scaling technique for eight-node solid-shell elements is pro-
posed with the purpose of increasing the element maximum stable time-step in explicit dynamics sim-
ulations, without affecting the global structural response in inertia dominated problems. The procedure
is conceptually based on a simple linear transformation of the element degrees of freedom. The scal-
ing procedure does not actually require to modify the element implementation, as it merely consists of
a modification of nodal accelerations. Therefore, the degrees of freedom transformation is not carried
out in practice. However, the transformation applied to both mass and stiffness matrices is conceptually
useful since it allows for a consistent analytical derivation of the element maximum eigenfrequency and,
hence, to an accurate estimate of the critical time-step size.
17
 Figure 10: Cylindrical shell hinged on two sides: snapshots showing snap-through configurations.

Figure 11: Cylindrical shell hinged on two sides: a) energy time evolutions; b) evolutions of energy error (dotted curve) and
stability threshold (solid curve).

The analytical estimate has been obtained for a parallelepiped element. In the case of elements which
are only slightly distorted in the original mesh or which become so due to structural deformation, a simple
strategy for the estimate of a lower bound of the critical time-step has been proposed. The derivation of
an accurate estimate in the case of highly distorted elements has still to be studied and will be the object
of future work.
Numerical examples have shown that the proposed selective mass scaling procedure allows for gains
up to almost two orders of magnitude in the time-step size, with only marginal accuracy loss. The
application of the proposed technique to a wider range of problems will be considered in the future with
the purpose of identifying optimal values of the scaling factor.
Another interesting future application of mass scaling techniques concerns shell and solid-shell el-
ements with thickness behavior enhanced in different ways (see e.g. [30–32]). Also in these cases the
enhancements are expected to increase the maximum eigenfrequency, requiring ad hoc provisions for its
estimate and reduction.

Acknowledgments

The financial support of Tetra Pak Packaging Solutions is gratefully acknowledged.

18
6. Appendix 1

The computation of the element maximum eigenfrequency implies finding the highest root of a cubic
equation of the type

C3 η3 + C2 η2 + C1 η + C0 = 0 (32)
The highest root is obtained using the following procedure (see also Box 1).

• Eliminate the quadratic term through a linear transformation of η

1 C2
η=ξ− (33)
3 C3

• Substitute (33) in (32), obtaining

!2 !3
C1 1 C2 C0 2 C2 1 C1 C2
ξ + p ξ + q = 0,
3
p= − , q= + − (34)
C3 3 C3 C3 27 C3 3 C3 C3

• Carry out a nonlinear transformation of ξ to obtain a quadratic equation:

p1  2 p3
ξ =u− → u3 + q u3 − =0 (35)
3u 27
whose solutions are r 
q q 2 p3
u =− ±
3
+ (36)
2 2 27
• Equation (36) provides six roots which, substituted back in (35)1 , allow to compute the highest
root of (32), which is given by
r
p ϕ 1 C
2
η9 = 2 − cos − (37)
3 3 3 C3
where r
p3
ϕ = arccos (−q/2µ) , µ= − (38)
27

7. Appendix 2

Eigenmodes of Group 2 (See Figure 3a).

Due to symmetry requirements, the eigenvectors U of this Group must take the following form:
   


 0  
 

 −u x  

   
ave


 0 

 diff


 −u x



Ux   
 U x

 

0 u

 
 
 



 

 

 x 


   



 0 






 u x




   
0 −u

 
 
 

 ave  
 
 
 y 


 U 
 

 0


 

 u



diff y
U4,5,6 =  , ave
= , diff
=
 12×1   ave
   
U U U U (39)
    
diff y y
 U 0  uy 

 

 

 

 

 

   
12×1    
0  −uy 


 
 

 



 

 

 


u 0
   
z


 

 

 


   
−u 0

 
 
 

ave
 z  diff  
Uz  U

 
 
 

 z  




 u z









 0 



 −uz 
   0 
 


19
where all entries are expressed in terms of the three parameters u x , uy and uz only. Following the same
path of reasoning discussed for Group 2 in Section 3.1, one obtains:
h  i
u x = 24 γ η γ λ + 24 1 − γ2 − λ2
h    i  
uy = α η2 γ2 λ − η 24γ α γ2 + λ2 + 1 + λ2 + 242 λ 1 + γ2 + λ2 (40)
h  i
uz = 24 α η γ λ − 24 1 − γ + λ
2 2

where α is the mass scaling factor defined in (12). Replacing the above solution in the original system
(23)1 , many of the equations result identically satisfied, while the remaining set of equations reduces to
a single cubic equation to be solved with respect to η:
h  i h  i  
α2 γ2 λ2 η3 −24 α γ λ 2 + (1 + α) γ2 + λ2 η2 +242 1 + α γ2 + λ2 1 + γ2 + λ2 η−4·243 γ λ = 0 (41)

Eigenmodes of Group 4 (See Figure 3c).

Due to symmetry requirements, the eigenvector U must take the following form:

U10,11 = {−u x u x − u x u x | 0 0 0 0 | 0 0 0 0 | 0 0 0 0 | 0 0 0 0 | uz uz − uz − uz }T (42)

Following the same path of reasoning as above, using just one of the 24 equations, one has:

u x = 48 γ ν h i (43)
uz = η γ λ (1 − 2ν) − 24 2γ2 (1 − ν) + λ2 (1 − 2ν)

where η solves the following second order equation, with real roots:
h   i
η2 α γ2 λ2 (1 − 2ν) − η 24 γ λ 2 1 + α γ2 (1 − ν) + λ2 (1 + α) (1 − 2ν) +
h   i (44)
+242 4γ2 + 2λ2 1 + γ2 (1 − ν) + λ4 (1 − 2ν) = 0

i.e.: h   i2
∆ = 2 1 − α γ2 (1 − ν) − λ2 (α − 1) (1 − 2ν) + 16 α γ2 ν2
h   i √ (45)
2 1 + αγ2 (1 − ν) + λ2 (1 + α) (1 − 2ν) ± ∆
η10,11 = 12
αγλ (1 − 2ν)

Eigenmodes of Group 5 (See Figure 3d).

Due to symmetry requirements, the eigenvector U must take the following form:
n oT
U12,13 = 0 0 0 0 | − uy uy − uy uy | 0 0 0 0 | 0 0 0 0 | 0 0 0 0 | uz − uz − uz uz (46)

Following the same path of reasoning as above, using just one of the 24 equations, one has:

uy = 48λν h i (47)
uz = η γ λ (1 − 2ν) − 24 γ2 (1 − 2ν) + 2λ2 (1 − ν)

where η solves the following second order equation, with real roots:
h   i
η2 α γ2 λ2 (1 − 2ν) − η 24 γ λ 2 1 + α λ2 (1 − ν) + γ2 (1 + α) (1 − 2ν) +
h   i (48)
+242 4λ2 + 2γ2 1 + λ2 (1 − ν) + γ4 (1 − 2ν) = 0

20
i.e.: h   i2
∆ = 2 1 − α λ2 (1 − ν) − γ2 (α − 1) (1 − 2ν) + 16 α λ2 ν2
h   i √
2 1 + αλ2 (1 − ν) + γ2 (α + 1) (1 − 2ν) ± ∆ (49)
η12,13 = 12
α γ λ (1 − 2ν)

Eigenmodes of Group 6 (See Figure 3e).

Due to symmetry requirements, the eigenvector U must take the following form:
n oT
U14,15 = 0 0 0 0 | 0 0 0 0 | 0 0 0 0 | u x − u x − u x u x | uy uy − uy − uy | 0 0 0 0 (50)

Following the same path of reasoning as above, using just one of the 24 equations, one has:

u x = 48 γ λ ν h i (51)
uy = η α γ λ (1 − 2ν) − 24 2γ2 (1 − ν) + (1 − 2ν)

where η solves the following second order equation, with real roots:
h  i
η2 α2 γ2 λ2 (1 − 2ν) − η 48 α γ λ γ2 + λ2 (1 − ν) + (1 − 2ν) +
h   i (52)
+242 4γ2 λ2 + 2 γ2 + λ2 (1 − ν) + (1 − 2ν) = 0

i.e.:  2
∆ = γ2 − λ2 (1 − ν)2 + 4γ2 λ2 ν2
  √
γ2 + λ2 (1 − ν) + (1 − 2ν) ± ∆ (53)
η14,15 = 24
α γ λ (1 − 2ν)

Eigenmodes of Group 7 (See Figure 3f).

Due to symmetry requirements, the eigenvector U must take the following form:
n oT
U16,17 = −u x − u x u x u x | − uy uy uy − uy | 0 0 0 0 | 0 0 0 0 | 0 0 0 0 | 0 0 0 0 (54)

Following the same path of reasoning as above, using just one of the 24 equations, one has:

u x = 72 γ
(55)
uy = η γ − 72 λ

where η solves the following quadratic equation:

h  i
η η γ λ − 72 γ2 + λ2 = 0 (56)

i.e.: 

 0
η16,17 = γ λ
 !
(57)

 72
 +
λ γ


The zero eigenvalue corresponds to one of the three rigid body rotation eigenmodes.

Eigenmodes of Group 8 (See Figure 3g).

Due to symmetry requirements, the eigenvector U must take the following form:
n oT
U18,19 = 0 0 0 0 | 0 0 0 0 | − uz − uz uz uz | 0 0 0 0 | uy uy uy uy | 0 0 0 0 (58)
21
Following the same path of reasoning as above, using just one of the 24 equations, one has:

uy = η γ − 72λ
(59)
uz = 72

where η solves the following quadratic equation:

h  i
η η α γ λ − 72 1 + αλ2 = 0 (60)

i.e.: 

 0
η18,19 = λ
 !
1 (61)

 72
 +
αγλ γ


The zero eigenvalue corresponds to the second of the three rigid body rotation eigenmodes.

Eigenmodes of Group 9 (See Figure 3h).

Due to symmetry requirements, the eigenvector U must take the following form:

U20,21 = {0 0 0 0 | 0 0 0 0 | − uz uz uz − uz | u x u x u x u x | 0 0 0 0 | 0 0 0 0}T

Following the same path of reasoning as above, using just one of the 24 equations, one has:

u x = η λ − 72γ
(62)
uz = 72

where η solves the following quadratic equation:

h  i
η η α λ γ − 72 1 + α γ2 = 0 (63)

i.e.: 

 0
η20,21 = γ
 !
1 (64)

 72
 +
αλγ λ


The zero eigenvalue corresponds to the last rigid body rotation eigenmode.

Eigenmodes of Group 10 (See Figure 3i).

Due to symmetry requirements, the eigenvector U must take the following form:

U22 = {0 0 0 0 | 0 0 0 0 | 0 0 0 0 | − u x u x − u x u x | 0 0 0 0 | 0 0 0 0}T (65)

Following the same path of reasoning as above, one immediately has that it can be assumed:

ux = 1 (66)

and η is given by the following expression:

λ γ 1−ν
!
8 1
η22 = + +2 (67)
α γλ γ λ 1 − 2ν

Eigenmodes of Group 11 (See Figure 3l).

22
 Due to symmetry requirements, the eigenvector U must take the following form:
n oT
U23 = 0 0 0 0 | 0 0 0 0 | 0 0 0 0 | 0 0 0 0 | − uy uy − uy uy | 0 0 0 0 (68)

Following the same path of reasoning as above, one immediately has that it can be assumed:

uy = 1 (69)

and η is given by the following expression:

γ λ 1−ν
!
8 1
η23 = + +2 (70)
α γλ λ γ 1 − 2ν

Eigenmodes of Group 12 (See Figure 3m).

Due to symmetry requirements, the eigenvector U must take the following form:

U24 = {0 0 0 0 | 0 0 0 0 | 0 0 0 0 | 0 0 0 0 | 0 0 0 0 | − uz uz − uz uz }T (71)

Following the same path of reasoning as above, one immediately has that it can be assumed:

uz = 1 (72)

and η is given by the following expression:

8 γ λ 1 1−ν
!
η24 = + +2 (73)
α λ γ γ λ 1 − 2ν

References
[1] Buechter, N., Ramm, E.. Shell theory versus degeneration - a comparison in large rotation finite element analysis.
International Journal for Numerical Methods in Engineering 1992;34(1):39–59.
[2] Bathe, K.J., Dvorkin, E.N.. A four-node plate bending element based on Mindlin/Reissner plate theory and a mixed
interpolation. International Journal for Numerical Methods in Engineering 1985;21(2):367–383.
[3] Simo, J.C., Fox, D.D.. On a stress resultant geometrically exact shell model. part i: Formulation and optimal parametriza-
tion. Computer Methods in Applied Mechanics and Engineering 1989;72(3):267–304.
[4] Chapelle, D., Bathe, K.J.. The Finite Element Analysis of Shells - Fundamentals. Springer; 2003. ISBN 3-540-41339-1.
[5] Parisch, H.. A continuum-based shell theory for non-linear applications. International Journal for Numerical Methods in
Engineering 1995;38(11):1855–1883.
[6] Hauptmann, R., Schweizerhof, K.. A systematic development of ’solid-shell’ element formulations for linear and
non-linear analyses employing only displacement degrees of freedom. International Journal for Numerical Methods in
Engineering 1998;42(1):49–69.
[7] Vu-Quoc, L., Tan, X.G.. Optimal solid shells for non-linear analyses of multilayer composites. i. statics. Computer
Methods in Applied Mechanics and Engineering 2003;192(9-10):975–1016.
[8] Ibrahimbegovic, A., Brank, B., Courtois, P.. Stress resultant geometrically exact form of classical shell model and
vector-like parameterization of constrained finite rotations. International Journal for Numerical Methods in Engineering
2001;52(11):1235–1252.
[9] Cardoso, R.P.R., Yoon, J.W., Mahardika, M., Choudhry, S., Alves de Sousa, R.J., Fontes Valente, R.A.. Enhanced
assumed strain (EAS) and assumed natural strain (ANS) methods for one-point quadrature solid-shell elements. Interna-
tional Journal for Numerical Methods in Engineering 2008;75(2):156–187.
[10] Schwarze, M., Reese, S.. A reduced integration solid-shell finite element based on the EAS and the ANS concept:
Geometrically linear problems. International Journal for Numerical Methods in Engineering 2009;80(10):1322–1355.
[11] Abed-Meraim, F., Combescure, A.. An improved assumed strain solid-shell element formulation with physical stabi-
lization for geometric non-linear applications and elastic-plastic stability analysis. International Journal for Numerical
Methods in Engineering 2009;80(13):1640–1686.
[12] Schwarze, M., Reese, S.. A reduced integration solid-shell finite element based on the EAS and the ANS concept: Large
deformation problems. International Journal for Numerical Methods in Engineering 2011;85(3):289–329.
[13] Wall, W.A., Gee, M., Ramm, E.. The challenge of a three-dimensional shell formulation: the conditioning problem. In:
Proc. IASS-IACM 2000 Fourth International Colloquium on Computation for Shells & Spatial Structures. 2000, p. 1–21.
23
 [14] Gee, M., Ramm, E., Wall, W.. Parallel multilevel solution of nonlinear shell structures. Computer Methods in Applied
Mechanics and Engineering 2005;194(21-24):2513–2533.
[15] Klöppel, T., Gee, M.W., Wall, W.A.. A scaled thickness conditioning for solid- and solid-shell discretizations of
thin-walled structures. Computer Methods in Applied Mechanics and Engineering 2011;200(9-12):1301–1310.
[16] Macek, R.W., Aubert, B.H.. A mass penalty technique to control the critical time increment in explicit dynamic finite
element analyses. Earthquake Engineering and Structural Dynamics 1995;24(10):1315–1331.
[17] Olovsson, L., Simonsson, K., Unosson, M.. Selective mass scaling for explicit finite element analyses. International
Journal for Numerical Methods in Engineering 2005;63(10):1436–1445.
[18] Olovsson, L., Simonsson, K.. Iterative solution technique in selective mass scaling. Communications in Numerical
Methods in Engineering 2006;22(1):77–82.
[19] Hetherington, J., Askes, H.. Penalty methods for time domain computational dynamics based on positive and negative
inertia. Computers & Structures 2009;87(23-24):1474–1482.
[20] Askes, H., Nguyen, D., Tyas, A.. Increasing the critical time step: micro-inertia, inertia penalties and mass scaling.
Computational Mechanics 2011;47(6):657–667.
[21] Key, S.W., Beisinger, Z.E.. The transient dynamic analysis of thin shells by the finite element method. In: Proc. of the
Third Conference on Matrix Methods in Structural Mechanics. 1971,.
[22] Gershgorin, S.. Ueber die abgrenzung der eigenwerte einer matrix. Izv Akad Nauk SSSR Ser Mat 1931;1:749–754.
[23] Flanagan, D.P., Belytschko, T.. Eigenvalues and Stable Time Steps for the Uniform Strain Hexahedron and Quadrilateral.
Journal of Applied Mechanics 1984;51(1):35–40.
[24] Hallquist, J.O.. LS-Dyna Theoretical Manual. Livermore Software Technology Corporation; 2002.
[25] Ling, X., Cherukuri, H.P.. Stability analysis of an explicit finite element scheme for plane wave motions in elastic solids.
Computational Mechanics 2002;29(4):430–440.
[26] Comini, G., Manzan, M.. Stability characteristics of time integration schemes for finite elementsolutions of conduction-
type problems. International Journal of Numerical Methods for Heat & Fluid Flow 1994;4(2):131–142.
[27] Hinton, E., Rock, T., Zienkiewicz, O.C.. A note on mass lumping and related processes in the finite element method.
Earthquake Engineering and Structural Dynamics 1976;4(3):245–249.
[28] Olovsson, L., Unosson, M., Simonsson, K.. Selective mass scaling for thin walled structures modeled with tri-linear
solid elements. Computational Mechanics 2004;34(2):134–136.
[29] Büchter, N., Ramm, E.. 3d–Extension on nonlinear shell eqautions based on the enhanced assumed strain concept. In:
Hirsch, C., editor. Computational Mechanics in Applied Sciences. Elsevier; 1992, p. 55–66.
[30] Bischoff, M., Ramm, E.. On the physical significance of higher order kinematic and static variables in a three-
dimensional shell formulation. International Journal of Solids and Structures 2000;37(46-47):6933–6960.
[31] Brank, B., Korelc, J., Ibrahimbegović, A.. Nonlinear shell problem formulation accounting for through-the-thickness
stretching and its finite element implementation. Computers & Structures 2002;80(9-10):699–717. doi:10.1016/S0045-
7949(02)00042-1.
[32] Kim, D., Bathe, K.. A 4-node 3D-shell element to model shell surface tractions and incompressible behavior. Computers
& Structures 2008;86(21-22):2027–2041.
[33] Belytschko, T., Liu, W.K., Moran, B.. Nonlinear finite elements for continua and structures. Wiley; 2000.

24

View publication stats