Presented

by
Dr Ceri
Computational Methods -
Morris Week 8

Discretisation
 Discretisation
• Process by which a differential or integral equation is
approximated by algebraic expressions to provide values at
a number of discrete points in the flow domain

• While an analytical solution provides a description over the

whole flow domain, a numerical CFD solution only provides
numerical values for the solution at discrete points in the
domain called grid points, mesh points, or nodes.

• Computers cannot solve differential (or integral) equations

unless they are put in a discretised form. This means with
CFD the PDEs are changed into algebraic equations
Discretisation Schemes
• Finite Difference
• Finite Volume
• Finite Element

• Less common: spectral methods, boundary element

methods and particle (vortex) methods

• Simulate finite problem of an infinite physical world;

leads to loss of information
The Finite Difference Method
• The FDM is based on the definition of derivatives
and on Taylor series expansions

• It is the easiest method to understand and apply

• It requires the use of a structured grid

(hexahedral or quadrilateral elements only)
 Taylor Series

𝜕𝑓
𝑓 𝑥 + ∆𝑥
𝜕𝑥
𝜕𝑓 𝜕 𝑓 ∆𝑥
𝑓 𝑥 + ∆𝑥 +
𝜕𝑥 𝜕𝑥 2!
𝜕𝑓 𝜕 𝑓 ∆𝑥 𝜕 𝑓 ∆𝑥
𝑓 𝑥 + ∆𝑥 + +
𝜕𝑥 𝜕𝑥 2! 𝜕𝑥 3!
𝑓(𝑥)

𝑓 𝑥

𝑥 𝑥+∆𝑥
 Taylor Series

𝑓 0.22 ≈0.9899 0.775% error

𝑓 𝑥 = sin 2𝜋𝑥

𝑓 0.22 = 0.9823
𝑓 0.22 ≈0.9824 0.001% error
𝑓(𝑥)

𝑓 0.2 = 0.9511

𝑓 0.22 ≈ 0.9511 3.176% error

𝑥 𝑥 + ∆𝑥
 Finite difference
• Let us consider the terms in the Navier-Stokes equation
for incompressible flow in the x direction:

• Question: How can we approximate the partial

derivatives above?
• Answer: using the finite difference method. This
depends on the definition of derivatives and Taylor
series expansion.
u
• Let us see how the term may be approximated using
x
the FDM.
 First Order Forward Difference
Taylor Series Expansion:

, ,
, , ,

Rearrange for :
,
𝑖, 𝑗 𝑖 + 1, 𝑗
, ,

, , ,

Finite Difference Truncation Error

Representation

, , , ,

, ,

First Order
Forward difference
 First Order Backward Difference
Taylor Series Expansion:

, ,
, , ,

Rearrange for :
,
, , 𝑖 − 1, 𝑗 𝑖, 𝑗

, , ,

Finite Difference Truncation Error

Representation

, , , ,

, ,

First Order
Backward difference
Second Order Central Difference
Subtract Taylor Series expansion for , from expansion for , :

, ,
, ,

Rearrange for :
,
, , 𝑖 − 1, 𝑗 𝑖 + 1, 𝑗

, ,

Finite Difference Truncation Error

Representation

, , , ,

, ,
Second Order
Central difference
 Second Order Central Second Difference

Navier-Stokes equation in x direction:

Summing Taylor Series expansions for , and , :

, , ,
, ,

Rearrange for :
,
𝑖 − 1, 𝑗 𝑖, 𝑗 𝑖 + 1, 𝑗
, , ,

,
Second Order
Central Second Difference
 Finite Difference Expressions
𝜕𝑢 𝑢 ,−𝑢, 𝜕𝑢 𝑢, −𝑢 , 𝜕𝑢 𝑢 , −𝑢 , 𝜕 𝑢 𝑢 , − 2𝑢 , + 𝑢 ,
= = = =
𝜕𝑥 ∆𝑥 𝜕𝑥 ∆𝑥 𝜕𝑥 ,
2∆𝑥 𝜕𝑥 (∆𝑥)
, , ,

𝑖, 𝑗 𝑖 + 1, 𝑗
𝑖 − 1, 𝑗

1st order 1st order 2nd order 2nd order

forward difference backward difference central difference central 2nd difference
wrt x wrt x wrt x wrt x

𝜕𝑢 𝑢, −𝑢, 𝜕𝑢 𝑢, −𝑢, 𝜕𝑢 𝑢, −𝑢, 𝜕 𝑢 𝑢, − 2𝑢 , + 𝑢 ,

= = = =
𝜕𝑦 ∆𝑦 𝜕𝑦 ∆𝑦 𝜕𝑦 ,
2∆𝑦 𝜕𝑦 (∆𝑦)
, , ,
𝑖, 𝑗 + 1

𝑖, 𝑗 − 1
1st
order 1st
order 2nd order 2nd order
forward difference backward difference central difference central 2nd difference
wrt y wrt y wrt y wrt y
Higher Order Accuracy Finite Differences
E.g.
, , , , ,

,
Fourth Order
Central difference

• Can be derived by repeated application of Taylor’s series expanded about

grid points (i+1,j), (i,j), (i-1,j) 𝑖 − 1, 𝑗 𝑖 + 1, 𝑗
• Involves information at 5 grid points 𝑖 − 2, 𝑗 𝑖, 𝑗 𝑖 + 2, 𝑗

• Require more computer time since more grid points required

• May require fewer total grid points to achieve comparable accuracy
• May result in higher quality solution
Difference Equations
• A partial differential equation where all of the partial
derivatives are replaced by finite-difference quotients

• E.g. 2D Navier-Stokes
steady momeequation in x direction:

• Difference Equation using central differencing:

, , , ,

, , , , , ,
 Example 1: 1D linear Convection
• Consider the equation
u u
c 0
t x
c = is the transport velocity, u = u (x,t) is a transported quantity
Since c = constant this a linear problem.
u u
• There are two partial derivatives: t and x so we need to discretise in
time t and space x: time-space discretisation.

• Let i be the index of the grid in x spatial direction

• Let n be the index of the grid in t
• Use a Finite Difference scheme such that:
• Forward differencing in time:
Superscript denotes change in time
n 1
u u  u n
 i i
Subscript denotes change in spatial
t t position

• Backward differencing in space

u uin  uin1

x x
The discretised form of the equation is then

uin 1  uin uin  uin1

c 0
t x

n and n+1 are two consecutive steps in time while i-1 and i are two
neighbouring spatial points.
 Example 2: 1D Diffusion equation
u  2u
• Consider the PDE t   x 2
• Use Forward difference in time t, with index n
• Use central difference for space x, with index i
• The discretised form is uin1  uin  uin1  2uin  uin1 
   
t  x 2

• Solving for u n 1
i

 t
u n 1
u 
n
u n
 2uin  uin1 
i i
x  2 i 1
 FD Method: Observations

• FDM is the oldest and easiest to apply, but its reliance on

a structured mesh to discretise the system of PDEs makes
it a method hard to apply to complex geometries in
multiple dimensions.

• This has motivated the development of more generalised

methods: finite element and finite volume techniques
 Finite Volume Method (FVM)

• With FDM, the governing equations (in differential form) are

discretised, i.e. changed into a set of algebraic (arithmetic)
equations.

• FVM, however, uses a volume integral formulation of the

problem with a finite set of volumes to discretise the equations.

• Applicable on general unstructured mesh (can be used on

structured meshes too)

• FVM is most widely used (~85%) in engineering simulations

 Finite Volume Method (FVM)
• Domain is discretised into a finite set
of control volumes (CV)

• Integral equations are transformed

into a set of algebraic (arithmetic)
equations, solved numerically by a
computer

• On each of the control volumes the

fluid equations based on the The fluid region of pipe flow
is discretised into a finite
conservation of mass, momentum
number of control volumes
and energy are then solved

• ANSYS Fluent CFD solvers are based

on the finite volume (FV) method.
Explicit Vs Implicit
• Difference equation • Dependent variables
with only 1 unknown based on several
for each variable unknowns

• Coupled sets of
equations

• Matrix or iterative
technique needed
Explicit Vs Implicit
• Simpler to set up and • Can use much larger values
program of Δt – fewer time steps
required

• In some cases Δt must be • More complicated to set

very small to maintain up and program
stability – can result in long • Computer time per time
running times step is much larger
• Large Δt results in large
truncation error for
transient simulations
Presented
by
Dr Ceri
Computational Methods -
Morris Week 8

Iterative Solvers
Iterative Solvers
• Consider simultaneous 4x + 2y + 3z = 8
equations shown: 3x - 5y + 2z = -14
-2x + 3y + 8z = 27

• Rearrange first equation to x = (8 -2y - 3z) / 4

give an expression for x
• Rearrange second equation y = (-14 - 3x - 2z) / -5
to give an expression for y
• Rearrange third equation to
give an expression for z z = (27 + 2x - 3y) / 8
Jacobi Method
Iteration 1
• First set y and z to 0 and solve for x
• Next set x and z to 0 and solve for y
• Then set x and y to 0 and solve for z

Iteration 2
• Take the y and z values from iteration 1 and solve for x
• Take the x and z values from iteration 1 and solve for y
• Take the y and z values from iteration 1 and solve for z

Iteration 3 – Take values from iteration 2 and continue…..

 Iterative Solvers - Jacobi
0 1 2 3 4 5 6 7 8 9 10
X 0 2.000 -1.931 -2.794 -0.050 0.033 -1.819 -1.546 -0.355 -0.749 -1.476

Y 0 2.800 5.350 2.771 1.478 3.425 3.943 2.548 2.449 3.400 3.297

Z 0 3.375 2.825 0.886 1.637 2.808 2.099 1.442 2.033 2.368 1.913

2
Value

X
0 Y
-2 Z

-4
0 5 10 15 20 25 30
Iteration
Residuals
• Move all terms in each expression to the left hand side
• The right hand side should then equal zero
• Enter the values calculated for each iteration to give the
residual for that iteration
• Residuals show how far away the estimated values are
from perfectly satisfying the equation
 Residuals - Jacobi
0 1 2 3 4 5 6 7 8 9 10
X 0.00 2.00 -1.93
-8.00 15.73 3.45
Y 0.00 2.80 5.35
14.00 12.75 -12.89
Z 0.00 3.38 2.83
-27.00 4.40 15.51
30

25

20 X
Residual

15 Y
10 Z

0
0 5 10 15 20 25 30
Iteration
 Iterative Solvers – Gauss-Seidel
0 1 2 3 4 5 6 7 8 9 10
X 0.00 2.00 -1.78

Y 0.00 4.00 2.68

Z 0.00 2.38 1.92

5
4
3
2
Value

1 X
0 Y
-1 Z

-2
-3
0 2 4 6 8 10
Iteration
Residuals – Gauss-Seidel
0 1 2 3 4 5 6 7 8 9 10
X 0.00 2.00 -1.78
-8.00 15.13
Y 0.00 4.00 2.68
14.00 4.75
Z 0.00 2.38 1.92
-27.00 0.00
30
25
20 X
Residual

15 Y
10
Z
5
0
-5 0 2 4 6 8 10
Iteration
Initial Values
• Possible to use better estimates as initial values
instead of setting all variables to 0
• Reduce the number of iterations needed to
achieve a stable solution
• Common to set all values to those at a boundary,
e.g. inlet
Iterative Solvers
• These examples used very simple equations.
Imagine how long it would take to find a
solution to the Navier-Stokes equations for
several thousand or million cells

• Hence we need to use computers!

 Relaxation Factors
• It is common to apply relaxation as follows:
 new , used
P   old
P  U ( new , predicted
P  old
P )
• Here U is the relaxation factor:
o U < 1 is underrelaxation. This may slow down speed of
convergence but increases the stability of the calculation, i.e.
it decreases the possibility of divergence or oscillations in
the solutions.
o U = 1 corresponds to no relaxation. One uses the predicted
value of the variable.
o U > 1 is overrelaxation. It can sometimes be used to
accelerate convergence but will decrease the stability of the
calculation.
 Relaxation Factors

 new , used
P   old
P  U ( new , predicted
P  old
P )
• E.g. If velocity at a given point is calculated as 2 m/s after
10 iterations and the new, predicted value after iteration
11 is 2.4 m/s
• A U value of 0.5 will result in a new, used value for
iteration 11 of 2.2 m/s
• A U value of 0.75 will results in a new, used value for
iteration 11 of 2.3 m/s
• A U value of 1.5 will result in a new, used value for
iteration 11 of 2.6 m/s
 Can you?
• Explain the difference between Explicit & Implicit Methods
o Explicit – 1 unknown per variable, simple, need small Δt for
stability
o Implicit - variables based on several unknowns, matrix/iterative
approach needed, more complex but stable for larger Δt

• Solve simple equations using Jacobi and Gauss-Seidel

methods and explain why Gauss-Seidel usually converges
more quickly.

• Explain why relaxation factors are used and the effect of

changing them