Introduction to computational

grid and time step

Computational Fluid Dynamics

* Includes some topics of numerical method class
* : Skipped in the exam.
 Sources of error in CFD ( simulation )
• Human error . mastake
– ex) wrong viscosity value
– Solution : 확인 & 확인

• Modeling error
– Error in the differential equation
4 deat
– ex)  = , constant density for aero-acoustics, etc. gas Frealgas
– Can be human error
– Solution : 공부 & 확인

• Discretization error (truncation error)

– Error from discretization method in terms of Taylor series expansion
– Solution : Reduced by using a smaller grid spacing

• Round-off error
– Machine epsilon : The smallest real number represented by computer
– Below this, the numeric information is lost
– Related to computer memory architecture
• The smallest number expressed by 4-byte real number : 10~
– Solution : Use 8-byte real number (long float in C, r8 in Star-CCM+)
 Sources of error in CFD
• Convergence error
– Error in the numerical solution
– Only for an iterative method for solving matrix
• Simulation은 기본적으로 연립 방정식을 푸는 것
– 연립 방정식은 기본적으로 matrix식 Ax=b로 표현 가능
• Direct method * 5M .
Unknown

F
.

 =   * 25o16 . mafrixelements

다
• Iterative method
* 4byteX 25시 lb
. = lol4 bytes
– 대부분의 경우에 해당

 =   +  Convergence error

– Solution : Convergence
…
(residual) criterion (수렴 조건)을 낮춤
x Time 가
 Why computational grid is necessary
• Governing equation for heat conduction
∇  = 

– Poisson equation
– Categorized as an elliptic PDE
– q is the heat source/sink

• In 1-D
  
In Math class = () =  +   + 
  2

In CFD class ℎ ℎ ℎ ℎ

T i−2 T i −1 Ti T i +1 T i+2

– h is the grid spacing

 Why computational grid is necessary
• In 1-D
h h h h
T i−2 T i −1 Ti T i +1 T i+2

  − 2 + 

= =  , " = 1,2, ⋯ , $
  ℎ
– Governing differential equation => a series of algebraic equations

• In 3-D
 Computational grid (mesh) in 2-D or 3-D
• Terminology for computational grid
small C
O
V
node
.

y
/

Cell = control volume into which cell center

domain is broken up.

Ocoredge )
face

ㅇ
Node = grid point.
cell

Cell center = center of a cell. sface center 2D computational grid

O
Edge = boundary of a face.

ㅇ
Face = boundary of a cell.
ㅇ node

O ㅇ
edge
"
face cell

3D computational grid
 Computational grid in 2-D or 3-D
• Topology for computational mesh

hexahedron
… prism or wedge
~

6 5

polyhedron
ㅡ

(arbitrary)
tetrahedron
pyramid 다면체
F 5
 Computational grid in 2-D or 3-D
• Examples : hexahedral mesh, trimmerftrimmed
d
Otfine

ㅇ ㅇ

ㅇ
 Computational grid in 2-D or 3-D
• Examples : tetrahedral mesh
 Computational grid in 2-D or 3-D
• Examples : polyhedral mesh

O
fine L
 Computational grid in 2-D or 3-D
• Examples of mesh
– Tri/Tetrahedral mesh vs. Quad/Hexahedral mesh

[ *
• Tri/Tetrahedral mesh : Easier (faster) generation
• Quad/Hexahedral mesh : More accurate solution

.
다뀔 .. .
 Computational grid in 2-D or 3-D
• Examples : Hybrid mesh
– Types by region
tet mesh
Specific regions can be meshed with
different cell types.
hex mesh

Both efficiency and accuracy are enhanced

relative to a hexahedral or tetrahedral mesh
alone.

wedge mesh

Hybrid mesh for a valve port

 Computational grid in 2-D or 3-D
• Examples : Hybrid mesh
– Types by orientation
– Polyhedral - Generalized Cylinder Mesher - Prism Layer

Tvhexahedral
polyhedral >
 Computational grid in 2-D or 3-D
• Structured/Unstructured mesh
Single-Block Structured Mesh
- i,j,k indexing to locate neighboring cells + +
- Grid lines must pass all through domain & + +
여
→ Obviously can’t be used for very complicated geometries 딸

Multi-Block Structured Mesh

- Uses i,j,k indexing within each mesh block (ij )

- The grid can be made up of arbitrarily-connected blocks

→ More flexible than single block, but still limited

Unstructured Mesh
- The cells are arranged in an arbitrary fashion
- No i,j,k grid index, no constraints on cell layout (usually 1-D index)
→ There is some memory and CPU overhead for unstructured referencing 체 N ~
 Topics on computational grid
• How to make a computational grid and what is a good
grid?

텍
• How to convert a governing equation to a series of
algebraic equations?
– Discretization in space
– Finite difference method
1 –
–
Finite volume method
Finite element method

• What is the error from the approximation?

– Error evaluation using Taylor-series expansion
 Topics on computational time step
• How to convert a governing equation to a series of
algebraic equations?
– Discretization in time

• What is good discretization in time?

– On accuracy & error term
• Derived by Taylor series expansion
– On stability
• How large Δt can be used to solve a PDE or ODE?
– The larger, the better => we can finish a simulation earlier
• Stability analysis
디 stability
U
accuracy

( divergenceblow )
up
-
Taylor series expansion

Computational Fluid Dynamics

 fn
Papproximate
a .

Taylor series expansion

• An important theory to analyze numerical accuracy
FCextrapolation (guessing a function value) based on a
– An '

orlation
polynomial built using a passing point and derivative information
interl
)

at the point x
[ 보간)
present tirate
– Various forms
%
• Based on a linear polynomial (2 terms)
ㅡ

%
% ℎ = %(0) + ℎ + Ο ℎ

resent' slopetime
future
'

쁘
|

• Based on a quadratic polynomial (3 terms)

% ℎ   % 0 ℎ 

% ℎ = %(0) + ℎ + 
+ Ο ℎ'
 2 
• Based on a cubic polynomial (4 terms) iaterporlation
ㄱ
ㅡ

% ℎ   % ℎ'  ' % 요
% ℎ = %(0) + ℎ + 
+ '
+ Ο ℎ) X extraporlation
 2  6  고
O
difficult
• In most general form, (가장 일반적인 형태로) X
. + +
ℎ  %
% ℎ = %(0) + * +
+ Ο ℎ.
,! 
product
+/
 Error of Taylor series expansion
• From the general form of extrapolation
– Note that all derivatives are defined at x=0 ( Passing point)

.
ℎ+  + %
% ℎ = %(0) + * + 0
,!  +
+/

– Error dependency on $

• A reasonable guess on error

– As the number of terms increases ($ increases, 더 고차의 항이 들
어갈수록), the error of extrapolation decreases
• When $ becomes infinity, the extrapolation converges to the exact
value
6
이 ( 하고 ℎ+  + %
%12345 ℎ = %(0) + *
~ ,!  +
+/
 Error of Taylor series expansion
• Error for a finite number (=n) of terms
% ℎ   % ℎ'  ' % ℎ)  ) %
%12345 ℎ = %(0) + ℎ + + + +⋯
 2   6  ' 24  )
hN Ist error term
– For Taylor series with 2 terms ⇒ as ,

becomes dominan가
% ℎ   % ℎ'  ' % ℎ)  ) %
%12345 ℎ = %(0) + ℎ + + + +⋯
 2   6  ' 24  )
Extrapolation Error term exact guess

째
= -

• (Most dominant) Error is reduced to about ¼ when ℎ becomes ½

• Called as “second-order” error
prentotn leading
order error term
-

기
– For Taylor series with 3 terms
%12345 ℎ = %(0) + ℎ

+
2  
+O
% ℎ   % ℎ'  ' % ℎ)  ) %
+
6  ' 24  )
+⋯

Extrapolation Error term

• (Most dominant) Error is reduced to about 1/8 when ℎ becomes ½
• Called as “third-order” error 3 H - order
~
accuracy
( exponet fh ) .
 Error of Taylor series expansion
• Error for a finite number (=n) of terms
…

– In the most general form

. 6
ℎ+ + %
ℎ+  + %
%12345 ℎ = %(0) + * +* leading ordererror : A - th order
,!  + ,!  + -
>

+/ +/.

Extrapolation Error term

• Error term is reduced to about 1/2n when ℎ becomes ½

• Called as “(n)th-order” error term
• Called as “(n)th-order” accuracy

• From now on, Taylor series expansion is written as

.

% ℎ = %(0) + *
ℎ+  + %
,! 
+/
+ + Ο ℎ. O
Extrapolation Error term
Numerical differentiation using Taylor series expansion
• Accuracy analysis of numerical derivatives 00
%(0) %(ℎ) − %(−ℎ) 뉴
=
 2ℎ %(−2ℎ) %(−ℎ) %(0) %(ℎ) %(2ℎ)

• Error estimation using Taylor series expansion

% ℎ   % ℎ'  ' % 스⑤ ( O ( h 4
④
d( fhd '. f Ihol "이
"

% ℎ = %(0) + ℎ + + + Ο ℎ) (f 6
표
= )

 2   6  '

% ℎ   % ℎ'  ' %
% −ℎ = %(0) ㅇ
−ℎ +  − ' + Ο ℎ)
 2  6 
.

%(ℎ) − %(−ℎ)
=
1
%(0) + ℎ
%

+
ℎ   %
2  
+
ℎ'  ' %
6 
% ℎ   % ℎ'  ' %
'
+ Ο ℎ) 따빠
E"
5
2ℎ 2ℎ )
없"
"
x
− %(0) − ℎ + − + Ο ℎ
 2   6  '
num ,

1 % ℎ'  ' % )
% ℎ  ' % '
derivative = 2ℎ + + Ο ℎ = + + Ο ℎ
2ℎ  3  '  6  '
.
guess (exact) Derivative Second-order error term
ㅡ
 Error estimation using Taylor series expansion
• Accuracy analysis of numerical derivatives
– Using Taylor series expansion
φ i −2 φ i −1 φi φ i +1 φ i+2
% % − %
= + Ο ℎ : second-order central difference
 2ℎ

% % − %
= +Ο ℎ : first-order forward difference
 ℎ

% % − %
= +Ο ℎ : first-order backward difference
 ℎ

% 3% − 4% + %

= + Ο ℎ : second-order backward difference
 2ℎ

% % − 8% + 8% − %

= + Ο ℎ) : fourth-order central difference
 12ℎ

  % % − 2% + %


= 
+ Ο ℎ
 ℎ
 Summary on Taylor series expansion
• Taylor series expansion is an extrapolation based on a
value and derivatives at a point
passing Point
• Size of the error (i.e. accuracy) is estimated based on the
exponent of the distance ℎ
– Second-order, third-order, nth-order
~ ㅡ

• Error of numerical derivatives (and almost everything in

1-D) can be evaluated using Taylor series expansion

• The size of error is affected by 3 factors

% ℎ   %
% ℎ = %(0) + ℎ + 
+ Ο ℎ'
 2 
Second-order error term
– Number of terms included in
extrapolation
( – ℎ : grid spacing
– Size of derivatives
Computational grid and
truncation error

Computational Fluid Dynamics

 OA D

Finite difference method (FDM)⇒ 쟬

• Taylor series expansion to approximate derivative
operators
)
; = ; + ℎ
; ℎ    ; ℎ '  ' ; ℎ )  ) ;

+
2   +
6  ' +
24  ) + Ο ℎ< e× y
; ; − ;
= + Ο ℎ FDM
 2ℎ

σC ;
>
+
;

=0 ⇒
;
>
+
; − ;
2ℎ

– Derivative of any order can be derived

=0 ⇒
; . − ; .
Δ>
+

음.
; − ;
2ℎ

udexofgrid
=0

n. indexof fimestep
• The size of error is affected by 3 factors
; − ; ; ℎ  ' ;
= + '
+ Ο ℎ'
2ℎ  6 

– Accuracy ~
Second-order error term
Hofterms
– ℎ : grid spacing
– Size of derivatives
~
 Conditions for a good grid in 1-D
• Good grid is one that minimizes the error
; − ; ; ℎ  ' ; '
= + + Ο ℎ
2ℎ  6  '
Error term
– Smaller ℎ => error term decreases
– Smaller size of derivatives => error term decreases D
• In order to reduce the error, small ℎ is necessary in the region where
the size of derivatives is large
• So, non-uniform ℎ may be desirable

– In 1-D, different grids can be generated by modifying ℎ (grid

spacings) 짧기
짧고
hi −2 hi −1 hi hi +1
T i−2 T i −1 Ti T i +1 T i+2
hp rforfast
.
.
computation
hN tfor
Cefficiency )
smallererrorl
 Conditions for a good grid in 1-D
• Good grid is one that minimizes the error
 ℎ     ℎ '  ' 
 =  + ℎ + + + Ο ℎ)
hi−2 hi −1 hi hi +1  2   6  '

 '
T i−2 T i −1 Ti T i +1 T i+2  ℎ    ℎ  ' 
 =  − ℎ + − + Ο ℎ)
 2   6  '

'
 −   ℎ − ℎ    1 ℎ + ℎ'  ' 
= +  + ' + Ο ℎ'
ℎ + ℎ  2  6 ℎ + ℎ 

Error term – first-order

– (hi-hi-1) increases => first-order error term increases

– Expansion ratio = hi / hi-1
• Expansion ratio close to 1 => good grid
 Conditions for a good grid in 1-D
• Summary
– A good grid can reduce the error
– Smaller overall grid spacing
– Smaller grid spacing in the region where the size of derivatives
~

are large onovuniformgrid

– Expansion ratio close to 1
 Conditions for a good grid in 2-D or 3-D
• Good grid is one that minimizes the error
– Smaller overall grid spacing
• More accurate but more expensive
– Smaller grid spacing in the region where the size of derivatives
are large
– Expansion ratio (인접한 cell사이의 크기 ratio) close to 1
• Which grid is better?
– Case #1
Same as 1-D case
smooth change large jump in
in cell size cell size
– Case #2 : 프로펠러 날개 끝 쪽의 속도 contours 태며
타력
sma ~

derivative
.

II
flarge
cerivative
Conditions for a good grid in 2-D or 3-D
– Case #2 : 프로펠러 날개 끝 쪽의 속도 contours

• Left one is better in terms of the expansion ratio

• Right one aims to reduce the error by using smaller grid spacing h
 Conditions for a good grid in 2-D or 3-D
• Good grid is one that minimizes the error
– Skewness (based on the optimum angle) close to zero =
정각형
= max C − CDE5FGF
• The optimum angle is 60 deg. for a triangle, 90 deg. for a rectangle,
etc. (angle between edges for a regular polygon (정각형))
• Examples
C = 90 θ max
θ min
Skewness = 0 Skewness > 0
optimal (equilateral) cell

circumcircle

actual cell

• How does skewness affect the accuracy?

 Conditions for a good grid in 2-D or 3-D
• How skewness affects the error
– Analysis of the error term is possible using the multi-dimensional
Taylor series expansion
∂φi ∂φi ∆x12 ∂ 2φi ∆x1 ∆y1 ∂ 2φi ∆y12 ∂ 2φi
φ1 = φi + ∆x1
∂x
+ ∆y1
∂y
+
2 ∂x∂x
+
2 ∂x∂y
+
2 ∂y∂y
+ Ο h2( )

1 i 2 1 i 2

3 3

∂φi φ2 − φ1 ∂φi φ2 − φ1
∂x
=
h
+ Ο h1 ( ) = + Ο h0 ( )
∂x h
∂φi φ1 + φ2 − 2φ3 ∂φi φ1 + φ2 − 2φ3
= + Ο(h1 ) = + Ο h0 ( )
∂y 3h ∂y 3h

– For case with zero skewness, 1st order accuracy but 0th order
accuracy for skewed grid
 Conditions for a good grid in 2-D or 3-D
• Good grid is one that minimizes the error
– Aspect ratio (종횡비)
• Aspect ratio is the ratio of longest edge length to shortest edge
length
• Equal to 1 (ideal) for a regular polygon (정각형)

aspect ratio = 1 high- aspect- ratio quad

aspect ratio = 1 high- aspect- ratio triangle

Good Not good

• How does the aspect ratio affect the accuracy?

 Conditions for a good grid in 2-D or 3-D
• How aspect ratio affects the error
– Consider a 2-D heat conduction problem with a large aspect
ratio
i, j + 1

∂ 2T ∂ 2T i − 1, j i, j i + 1, j
∇ 2T = q 2
+ 2 =q
∂x ∂y
i, j − 1

Ti −1, j − 2Ti , j + Ti +1, j Ti , j −1 − 2Ti , j + Ti , j +1

h 2
( )
+ Ο h 2x +
h 2
( )
+ Ο h 2y = qi , j
x y

– If hx = 100hy
Ti −1, j − 2Ti , j + Ti +1, j + 10000Ti , j −1 − 20000Ti , j + 10000Ti , j +1
h2x
( ) ( )
+ Ο 10000 h 2y + Ο h 2y = qi , j

• Size of coefficient in x-direction is relatively so small that it can be

ignored (x 방향은 제외한 것과 비슷한 효과)
– Error size in x-direction can increase a lot
 Summary of good grid in 2-D or 3-D
• Smaller overall grid spacing

서 • Smaller grid spacing in the region where the size of

derivatives are large

• Expansion ratio close to 1

• Skewness close to zero

삶 [• Aspect ratio close to 1

• + ??
Basics to select time step size &
grid spacing

Computational Fluid Dynamics

 Discretization in time
• "Computational grid" in time
∆t n −2 ∆t n −1 ∆t n ∆t n +1
; ;
+ =0
>  f n −2 f n −1 fn f n +1 f n +2

– In an unsteady (transient) simulation, the solutions at discrete

time stamps are obtained

• Finite difference method in time

– Use the Taylor series expansion to approximate derivative
operators
;. Δ>    ;. Δ> '  ' ;. Δ> )  ) ;.
;. = ;. + Δ> +  + ' + ) + Ο Δ> <
> 2 > 6 > 24 >
;. ;. − ;.
⇒ = + Ο Δ> 
> Δ>

– Δ> : computational time step -< 71 ,

h
General rules for computational time step
* n , MH fime
• Question: How to choose Δ>?
.
step
Egrid pt - in fime )

• Rule #1. The smaller, the more accurate ∂f

=
f n +1 − f n ∆t 2 ∂ 3 f
−
24 ∂t 3
– We know that the error decreases as Δ> decreases ∂t ∆t
f n +1 − f n
– How small should it be to get a reasonable solution?=
∆t
+ Ο ∆t 2 ( )

• Rule #2. Based on the size of time derivative

– Similar to discretization in space, we need small Δ> only in the
regions where the derivative (in time) is large

f (t )

t
small ∆t is desirable large ∆t is fine
1
'
forsmaller error
forefficienct
General rules for computational time step
• One question: How to choose Δ>?
• Rule #3 Based on the time scale
– See the “time scale” of flow variables…

– If Δ> is similar to the spacing between black points, what we get

from a simulation can be just a steady solution!
– Use several Δ> spacings (>=8) per an important time scale
Determining time step using time scale
• Examples of Rule #3
– From the “time scale” of flow variables…

• Case #1. Acoustics (noise) problem

– Important frequency of a sound wave ranges
• 50~2000 Hz (1/sec)
– The time scale ranges
• 0.0005~0.02 (sec)
– If we want 8 intervals in a time scale…
• Δ> = 0.0005/8~0.02/8 (sec)
• We have to choose the smaller one for accuracy
• => Δ> = 0.0005/8

• For a flow physics, you need to know the physical time

scale
– Ex) Smallest scale of turbulence
 Setting time step using time scale
• Examples of Rule #3
– From the “time scale” of flow variables…

• Case #2. Let’s guess the “time scale”

– A flow around a circular cylinder
I = 1 J/L M=1 J

– Try to compute the time scale by simply using the dimensions

P$Q>ℎ J M
"J LNO  L = = =1 L
RO">S J/L I

– In CFD, the observed time scale=2.7(s)

– Not perfect but not that bad

 Setting time step using time scale
• Examples of Rule #3

• Case #3. Guess the “time scale”

– Transonic airplane (Ma~1) of length=100m & height=10m
– Speed of airplane = speed of sound : about 300 m/sec
– Then, the time scale (sec) is
• About 100 (m) / 300 (m/sec)
• About 10 (m) / 300 (m/sec)
•  = 0.03~0.3 (sec)
– If we want 8 intervals in a time scale…
• Δ> = 0.03/8~0.3/8 (sec)
• Choose the smaller Δ> …
 Setting time step using CFL
• How to choose a right time step Δ> in a simulation?
– For accuracy & stability

• Rule #4 Based on CFL (Courant-Friedrich-Lewy) number

– Discussed in numerical PDE part
– Definition of CFL number in 1-D
. velocitySizelat )
VΔ> *
aghid P 마
TUP = Xo기 .
grid
Δ size

– Meaning of CFL number

• Distance a particle travels in a time step (u Δt) / grid spacing (Δx)
• Number of grid points a particle travels in a time step

– Typically used range of CFL number

• Less than CFL=10
• Typically 0.1~5
 Summary : Basics rules on time step
• To finish a simulation earlier
– The larger, the better

• For accuracy
– Rule #1. The smaller, the better
– Rule #2. Based on the size of time derivative
• Similar to discretization in space, we need small Δt only in the
regions where the derivative (in time) is large
– Rule #3 Based on the time scale
• Use several Δ> spacings (>=8) per an important time scale
– Rule #4 Based on CFL number
• CFL=0.1~5
Summary: Basics rules on grid for accuracy
• Rule #1. The smaller, the more accurate
• Rule #2. Based on the size of spatial derivative
• Rule #3. Based on the scale in space
– Example: A flow around a circular cylinder
• Contours of the x-velocity

m 2 ogridy
55
7

Necessary grid spacing?

• Length scale in x : similar to the distance between red spots (high U)

– About 5 (m)
• If we want more than 20 intervals in a length scale
– Δx <= 0.25 (m)
Good computational grid for accuracy
• Smaller overall grid spacing

• Smaller grid spacing in the region where the size of

derivatives are large

• Expansion ratio close to 1

• Skewness close to zero

• Aspect ratio close to 1

• Resolving important physical scales

 사례 : 벽면 격자를 위한 prism layer
• Where do you need a smaller grid size?
. large
change ( derivafive ) 어 Wal normal

d
*
x
 사례 : 벽면 격자를 위한 prism layer
• Modes of prism layer (Star-CCM+)
– Stretch factor
– Wall thickness
– Thickness ratio
 Stretch factor
• Parameters
– Number of Prism Layers
– Prism Layer Stretching
– Prism Layer Total Thickness

• Ex) Number of Prism Layers 2, Stretching 1.5, Total

thickness 5mm
– First Layer thickness 2mm, Second Thickness 3mm

smallest '
 Wall thickness
• Surface의 첫번째 layer thickness를 우선 지정
– Properties 기반 나머지 layer들의 thickness를 자동 조정

• Parameters
– Number of Prism Layers
– Prism Layer Near Wall Thickness
– Prism Layer Total Thickness
 Thickness ratio
• Parameters
– Number of Prism Layers
– Prism Layer Total Thickness
– Prism Layer Thickness Ratio
• 첫번째와 마지막 layer thickness의 비율
• 중간 layer들의 thickness는 자동의 계산
 surface
(
사례 : Volumetric control of grid spacing
• 특정 영역(refinement zone)에서 격자 조밀도 변화
원래 격자

중요한
유동 특성이
나타나는 곳

Volume
설정

부분적으로 격자를
조밀하게 변화