em pretEeviyzon
CHAPTER
4
BASIC
ASPECTS
OF
DISCRETIZATION
Numerical precision is the very soul of science,
Sir D'Arcy Wentworth Thompson, Scottish biologist
and natural scientist, 1917
4.1 INTRODUCTION
The word) “discretization” requires some explanation. Obviously, it comes fom
ct
defined in The American Heritage Dictionary of the English Language
dividual; distinct; consisting of unconnected
distinct parts.” However, the word “discretization” cannot be found in the same
dictionary; it cannot be found in Webster's New World Dictionary either. The fact
that it does not appear in two of the most popular dictionaries of today implies, at
the very least, that it is a rather new and esoteric word. Indeed, it seems to be unique)
to the literature of numerical analysis, first being introduced in the German literature
in 1955 by W. R. Wasow, carried on by Ames in 1965 in his classic book on partial
differential equations (Ref. 24), and recently embraced by the
found in Refs. 13, 14, and 16. In essence, discretization is the process by which a
closed-form mathematical expression, such as a function or a differential or integral
equation involving functions, all of which are viewed as having an infinite
continuum of values throughout some domain, ‘a analogous
(but different) expressions which prescribe values at only ‘number of discrete
points or volumes in the domain. This may sound a bit mysterious, so let us
elaborate for the sake of clarity. Also, we will single out partial differential
equations for purposes of discussion. Therefore, the remainder of this introductory
section dwells on the meaning of “discretization.”
125126. aasic aspects oF piscreTiZaTION
of partial differential equations i closed-form
ich give the variation of the dependent va
throughout the domain. dn
contrast, numerical solutions can give answers at only
discrete points in the domain, called grid points. For example, consider Fig. 4.1,
Which shows @ section of a discrete grid in the xy plane. For convenience, let us
that the
-and that the s
ein ate ve
spacing in bh cretion, where is dierent value between ech sueesve
pairs of grid points, and similarly for Ay. However, the majority of CFD applications
olve numerical solutions on a grid which contains spacing in each
n, because this greatly simplifies the pr of th
storage space, and usually results in greater accuracy. ci t
‘(ave to occur in he GRIM 2°: as is frequently done in CFD, the numerical
calculations are carried out in a transformed computational space which has uniform
spacing in the transformed independent variables but which corresponds to
nonuniform spacing in the physical plane. These matters will be discussed in
detail in Chap. 5. In any event, we will assume uniform spacing in
cing fc
we will assume Av and Ay'to but qual Aye (We
should note that recent reseatch in CFD has focused on unstructured gtids, where
the grid points are placed in the flow field in a very irregular fashion; this is in
contrast to a structured grid which reflects’some type of consistent geometrical
regularity. Figure 4.1 is an example of a structured grid. Some aspects of
unstructured grids will be discussed in Chap. 5.)
Returning to Fig. 4.1, /hich fins in
ee anidjan inde ich y . Hence, if (i, f) is the
index for point P in Fig. 4.1, then the point immediately to the right of P is labeled
as (i + 1, J), the point immediately to the left is @ — 1, /), the point directly above is
(47+ Vy and the point directly below is (i,j = 1).
FIG. 4.1
Discrete grid points.iwrropuction 127
equations, or as the case may be, the Euler equations, as derived in Chap. 2. These
An analytical solution of these equations would
provide, in principle, closed-form expressions for u, v, p, p, etc., a8 functions of x
and yp which could be used to give values of the flow-field vatiables at any point we
wish to choose in the flow, ie., at any of the infinite number of (x, y) points in the
domain. On the other hand, if the partial derivatives in the governing equations are
replaced by approximate algebraic difference quotients (to be derived in the next
section), whete the algebraic difference quotients are ssed strictly in terms of
the flow-field variables at wo or more ofthe discete gra points shown in Fig. 4.1,
then the partial differential equations are totally replaced by a system of pay
equations which can bi for the vi
. Moreover, this method of discretization is called the
es. Finite-difference solutions are widely employed in CFD, and
hence much of this chapter will be devoted to matters concerning finite differences.
So this is what discretization means. Al of
discretization. The purpose of this chapter is to derive and discuss the more
common forms of discretization in use today for finite-difference applications. This
constitutes one of the three main headings in Fig. 4.2, which is the road map for this
Diseretization techniques
Finite Finite Finite
difference volume element
Basic derivations Basic derivations
of finite differences: of finite-volume
order of accuracy | | equations (Problem 4.7)
difference
equations:
truncation errors
‘Types of solutions
explicit and implicit
Stability analysis
FIG. 42
Road map for Chap. 4.128 Basic aspects OF DISCRETIZATION
chapter. The seeofd and thiflimain headings are labeled finite Volume and finite
respectively. Both fi been in
widespread use in Computational mechanics for years. H s
or finite-element methods in this book, mainly because of
The essential aspects of finite volume discretization are dealt with via
Problem 4.7 at the end of this chapter. It is important to note that CRD caf be)
a using any
fi 42.
Examining the road map in Fig. 4.2 further, the purpose of the present chapter
is to construct the basic discretization formulas for finite differences, while at the
same time addressing the order of accuracy of these formulas. The road map in Fig.
4.2 gives us our marching orders—let’s go to
4.2 INTRODUCTION TO FINITE
DIFFERENCES
Here, we are interested in feplacing a partial/detivative with/alSuitable algebraié
‘difference quotient, ic. a finite difference. Most common finite-difference repre-
\sentations of derivatives are based on s. For example,
referring to Fig. 4.1, component of velocity at point (i, /), then
the velocit
expanded about point
Wey = + (5), (4.1)
+(e
1
Equation (4.1) if () the number of |
terms aad Seopa AMAT TRAD
Example 4.1. Since some readers may not be totally comfortable with the concept of
a Taylor series, we will review some aspects in this example,
enmas an a ee 2 ——-. ith all derivatives
So crn Gnaljoat tr
oF pe es (ax? af (Axy"
Fort Ax) =flx) + 5 Ae + Qt ge gt
[Note in Eq. (E.1) that we continue (0 use the partial derivative nomenclature to be
consistent with Eq, (4.1), although for a function of one variable, the derivatives in Eq.
(E.1) are really ordinary derivatives.] The significance of Eq. (E.1) is diagramed in
Fig. E4.1 that we know the value of fat x (point 1 in Fig. E4.1); we want to
calculate the value of fat x + Ax (point 2 in Fig. E4.1) using Eq. (E.1). Examining the
right-hand side of Eq. (E4.1), we see that the first term, f(x), is not a good guess for
J (x + Ax), unless, of course, the function f (x) is a horizontal line between points 1
and 2. An improved guess is made by approximately accounting for the slope of the
curve at point 1, which is the role of the second term, O//0x Ax, in Eq. (E.1). To obtain
an even better estimate of fat x + Ax, the third term, 0°//0x" (Ax)"/2, is added, which
approximately accounts for the curvature between points 1 and 2. Im|genéfal, to obtain
(E1)INTRODUCTION To FINITE o1rFERENCES 129
Sorayese) + Lar + % oT
wont een ‘eocane
FIG. E4.1
Tlustration of behavior of the first
—— three terms in a Taylor series (for
ae Example 4.1)
uded. Indee
carried on the right-hand side. To examine some numbers, let
f(x) = sin 2m
Atx=02: f(x) =0.9511 (£2)
This exact value of f (0.2) corresponds to point 1 in Fig. E4.1, Now, let Ax = 0,02, We
wish to evaluate f(x + Ax) = (0.22), From Eq. (E.2), we have the exact value:
Atx=0.22: f(x) = 0.9823
This corresponds to point 2 in Fig. E4.1. Now, let us estimate f (0.22) using Eq. (E.1).
Using just the first term on the right-hand side of Eq, (E.1), we have
f (0.22) = f(0.2) =
This corresponds to point 3 in Fig. E4.1, The percentage error in this estimate is
[(0.9823 — 0.9511)/0.9823] x 100 = 3.176 percent. Using two terms in Eq, (E.1),
9511
Set Ax) = f(x) + Zar
F(0.22) = f(0.2) + 20s [27(0.2)](0.02)
70.9511 +0.388 = 0.9899
This corresponds to point 4 in Fig. E4.1, The percentage error in this estimate is
[0.9899 — 0.9823)/0.9823] x 100 = 0.775 percent. This is much closer than the
previous estimate, Finally, to obtain yet an even better estimate, let us use three terms
in Eq, (E.1).130. aasic aseacts oF oiscerrzation
OF (Ax)?
Oe 2
‘£(0.22) ~ f (0.2) + 2mc0s [2(0.2)|(0.02) — 4x” sin {2n(0.2)| ———
0.9511 + 0.0388 — 0.0075
= 0.9824
This corresponds to point 5 in Fig. E4.1. The percentage error in this estimate is
(0.9824 — 0.9823)/0.9823] x 100 = 0.01 percent. This is a very close estimate of
f(0.22) using just three terms in the Taylor series given by Eq. (E.1).
set dx) ste) + Laws BF
(oon
Let us now return to Eq. (4.1) and pursue our discussion of finite-difference
representations of derivatives. SOIR Eq. (4.1) for(OWOX)g), we dbtain
au) _iviy—my (Pu) Ax (Ow) (Ax)?
(&),, ‘Ar ar),,2 \as),, 6 * 62)
ee
Finite-
difference ‘Truncation error
representation
he sl AEA RST GES
The firsfiiterm/on the right side, namely, (+1) — uj, VAX, is a finite-
difference representation of the partial derivative. The remaining terms on the right
side constitute the truncation error, That is, if we wish to approximate the partial
derivative with the above algebraic finite-difference quotient,
Quy Misty = Mis
( ae py Hh (43)
then the truncation erfomin Eq. (4.2) tells us what is being neglected) in this
TE, (4.2) the lowesvorder term inthe
approximation. the
‘Ax to the first power; hence, the finite-difference expression in Eq. (4.3) is called
stordeaccurate, We can more formally write Eq, (4.2) as
OuN _ Mivy — Ms
@ , + O(Ax) (4.4)
In Eq, (4.4), the syiibolO(Ax) isa fo tion which ts
“emsororden Ax.” Equation (4.4) Stich
d. AS a result, the finite difference in Eq. (4.4) is called a forward)
Gj) is uses
difference. For this reason, we now i urate difference
representation for |, expressed by Eq. (4.4) as a first-order
forward difference, repeated below.INTRODUCTION TO FINITE OIFFERENCES 131.
Bix
aN
(&) aa” G3)
Mig Lj — Mig
De + O(Ax)
J Gs) 44,3)
Let us now waite a/Taylor series expansion for upaigy, expanded about Wj
moins (5) a9 +(
Pu
“Ge,
or
Ou Pu\ (Ax?
@Pu\ (Ax) (43)
- ) ce bx
Mv KN
(Golving, for (@ulORy, we obtain “ey
a = (itn) 3) 4413)
@),- ce 43 4 O(Ax) (4.6)
to the
rearward (or backward) difference.
sa result,
The information) used in forming the finite-difference quotient in Eq. (4.6) comes
from tele of i pin th se el ea
ite difference in Eq. (4.6) is called a
the
in the trunca-
tion error involves Ax to the first power. As a result, the finite difference in Eq. (4.6)
is called a
curacy is fot Sufficient To
construct a finite-difference quotient of second-order accuracy, simply subtract
Eq. (4.5) from Eq, (4.1):
Ou ®u\ (Ax)?
ons mas=2 ) as+2(F5) ay,
hy hy)
Equation (4.7) can be written as
(3) = MHI MAY O(Ayy?
i
C443) fe) G43)
(48)
Ox, 2Ax
The information used in forming the finite-difference quotient in Eq, (4.8) comes
from hoth sides Of the grid point located at (i,,); that is, it uses upynyjas well as
j=, Grid point (i, falls between the two adjacent grid points, Moreover, in the132 _aasic asrects oF piscaeniZarion
truncation error in Eq. (4.7), the lowest-order terms involves (Ax)’, which is second-
order accuracy. Hence, the finite-differenice quotient in Eq. (4.8) is called a second-
order central difference.
‘he derivatives arejobtainedinexactly the’ same
fashion. (See Prob. 4.1 and 4.2.) The results are directly analogous to the previous
equations for the x derivatives. They are:
ee + O(Ay) Forward difference (4.9)
(2) _ wee + 0(Ay) Rearward difference (4.10)
hy)
ara +O(Ay)? Central difference, (4.11)
Equations (4.4), (4.6), and (4.8) to (4.11) are examples of finite-difference
quotients for first partial derivatives. Is this all that we need for CFD?\Let us return
to Chap. 2 for a moment and tak€a@l0oK at the governi ion. If we
mily, the governing equations are the Euler
b summarized = 2.8.2 and expressed Eqs. (2.82) to (2.86). Note
that the highest-order derivatives which
such as those
expressed by s. (4.4), (4.6), and (4.8), at
s. On the if we are dealing with
the governing equations are the Navier-Stokes equations, summarized in
Sec. 2.8.1 and expressed by Eqs. (2.29), (2.50), (2.56), and (2.66). Note that the
“highest-order_ defitves “which appear in the Navier-Stokes equations (ate
flect
appears in Eq. (2.66). anded, these terms such second
partial derivatives as
just for example.
there is a r derivatives oes CFD. We can
such fini -xpression: ith a TAylORSEHGs analysis,
as follows.
Summing the Taylor series expansions given by Eqs. (4.1) and (4.5), we have
Pu tw (Ax)*
yt Maa = 2m, +(33) ay +(F ) +
Mra tinny = Why +55 ) i 2
Solving for (2’w/Ax),, ;,
Pu Uy, 2a + Mi Mi
us Ol 4.
(Fi) ot =
In Eq. (4.12), the first term on the right-hand side is a central finite difference for the
second derivative with respect to x evaluated at grid point (i, /); from the remaining
order-of-magnitude term, we see that this central difference is of second-orderINTRODUCTION TO FINITE DIFFERENCES 133
accuracy. An analogous expression can easily be obtained for the second derivative
with respect t
oH hai — Why + my
3) a Mast Das aad ae 4 0(Ay)? (4.13)
td
Equations (4.12) and (4.13) are examples of second-order Central second
differences.
the case of mixed, derivatives, such as PulOx’dy, appropriate finite-
difference quotients c: \d as follows. Differentiating Eq. (4.1) with respect
toy, we have
Ow ‘ou ou) ( Ou ) (Ax? ( Ou ) (Ax?
= +H Art + spoon
Bani 6), Gry hj \OP y),, 2 "Oe w/),, 6
(4.14)
Differentiating Eq. (4.5) with respect to y, we have
CB)... @), Cera), 8 ee)
atu \ (Ax)
(3 hace ° (4.15)
Subtracting Eq. (4.15) from Eq. (4.14) yields
Es (%) (2 , ( ou ) (Ax)*
> -(— =2(-—]} Art +
G ing \OPAina3 — \O8 Ov); ax Ay}, ; 6
SELLER, which is the mixed derivative for which We are seeking’a
finite-difference expression, the above equation yields
es) - Ooo OUP E (en),
In Ea 16), the i tm ihehandsde involves Ou, Heese at
gtid point (i + 1, j) and then at grid point (i ~ 1, j). Returning to the grid sketched
in Fig. 4.1, we can see that Gu/Oy at each of these two grid points can be replaced
with a second-order central difference patterned 11) but
. To be
“more specific, in Eq. (4.16) first replace (OwAy),.+1, ; with
a) Up jet shied +4
[| eae OG:
G Heh) aa ca
(4.16)134 Basic ASPECTS OF DISCRETIZATION Sea ae aS ag dt
and then replace (Jwdv);1,, with the analogous difference, (ne
a ny] a
vj ee an
In this fashion, Eq, (4.16) becomes
N a LUr—s—S—e—SOOCOC_ONsNN
Ox Ov), 4dx Ay (4.17)
2 + O[(ax)?, (Ay)"]
» fiafan a Ax
in Eq. (4.11) is of O(Ay), Hence, the in Eq. (4.17) must/be
Ofary, any) Equation (4.17) gives a second-order central difference for the
mixed derivative, (Pw/x Oy),
It is important to note that when the governing flow equations are used in the
form of Eq. (2.93), only first derivatives are needed, even for viscous flows. The
dependent variables being differentiated are U, F, G, and H in Eq, (2.93), and only as
first derivatives. Hence, these derivatives can be replaced with the appropriate finite-
difference expressions for first derivatives, such as Eqs. (4.4), (4.6), and (4.8) to
(4.11). In turn, some elements of F, G, and H involve viscous stresses, such as t,.,
‘Tx, and thermal conduction terms. These terms depend on velocity or temperature
gradients, which are also first derivatives. Hence, the finite-difference forms for first
derivatives can also be used for the viscous terms inside F, G, and H. In this fashion,
the need to use a finite-difference expression for second derivatives, such as Eqs.
(4.12), (4.13), and (4.17), is circumvented.
To this stage, we have derived a number of different forms of finite-difference
expressions for various partial derivatives. To help reinforce these finite differences
in your mind, the graphical concept of finite-différence modules is useful.
LoS eae eR Te et context of t
in Fig. 4.3. This figure is a
cea wellaiustatng on geil especie gd
of each finite diference
n hese participating
gtid points are shown by large filled circles connected by bold lines;
schematic isicalledwfinite-ifference module. The
jin
‘The-truncation @fror in Eq. (4.17) (Comes from Eq. (4.16), where the
“order ied term is of O(Ax), gn the fact that the o
similarly, a
(—2) beside a grid point connotes that twice the variable at that grid point is
subtracted in the formation of the finite-difference quotient. Compare the (+), (—),
and (—2) in the finite-difference modules with the corresponding formula for the
finite difference which appears to the left of each module in Fig. 4.3.
The finite-difference expressions derived in this section and displayed in Fig.
4.3 represent just the “tip of the iceberg.” Many other difference approximations
can be obtained for the same derivatives we treated above. In particular, moreINTRODUCTION To FINITE DivERENCES 135
Firstorder
forward
difference
with respect.
tox
raved uy Mad = ey Ar
marward (Zt) = Hatt co) +)
with respect
tox
Second-order
central
difference
with respect
tox
FIG. 43
Finite-difference expressions with their appropriate finite-difference modules.
accurate finite-difference quotients can be derived, exhibiting third-order accuracy,
fourth¢order accuracy, and more. Such higher-order-accurate difference quotients
generally involve informationyat more grid points than those we have derived. For
example, a fourth-order-accurate central finite-difference for P°u/dx" is
() tia tN — 30M + 16H Hs, Oat (48)
ar), 12(An)?
Note that information at five grid points is required to form this fourth-order finite
difference; compare this with Eq. (4.12), where (0°u/0x°), ; is represented in terms
of information at only three grid points, albeit with only second-order accuracy.
Equation (4.18) can be derived by repeated application of Taylor’s series expanded
about grid points (/ + 1, /), (i,j), and (i — 1, {); the details are considered in Prob.
4.5, We are simply emphasizing that an almost unlimited number of finite-difference
expressions can be derived with ever-increasing accuracy. In the past, second-order
accuracy has been considered sufficient for most CFD applications, so the types of
difference quotients we have derived in this section have been, by far, the most136 Basic aspects oF DISCRETIZATION
Second-order
central
difference
with respect
toy
Met = Ms
2ay 4
Secoid-order a,
ein 22st Ms
second (ap, = ay
difference
with respect
toy
Seoond-
order
central (24) Meyert wins tendt tess
mixed \3xay/, “ax Ay
| R
= Z
respect -
toxandy Gy EL FL reeled
FIG. 43. (continued)2D
ecw OTIW Ws (SO) D> “\L> \ deustore AW Cour vee
Q J ) : aw