Numerical Methods in Thermo-Fluid Dynamics I

Initial value problems: Time-Stepping

Dr.-Ing. Manuel Münsch

Friedrich-Alexander Universität Erlangen-Nürnberg
Lehrstuhl für Strömungsmechanik

Erlangen
 Erlangen

Schedule
No. Week(s) Practical Lecture Exercise
Monday: 10:00-11:30 Monday: 14:15-15:45 Wednesday: 10:15-11:45
Room: online Room: online Room: online
1. 02.11.20-06.11.20 Matlab Tutorial I Introduction and Navier-Stokes Eq. No Exercise
2. 09.11.20-13.11.20 Matlab Tutorial II; Computer Exercise 1 Simplification of NSE up to Boundary Layer Eq. Hand out Exercise 1
3. 16.11.20-20.11.20 Work on Computer Ex. 1 Initial Value Problems: Time Stepping (ODEs) Work on Exercise 1
4. 23.11.20-27.11.20 Work on Computer Ex. 1 Finite Difference Method 1 Presentation Ex. 1 &
Hand out Ex. 2
5. 30.11.20-04.12.20 Work on Computer Ex. 1 Finite Difference Method 2 Work on Ex. 2
6. 07.12.20-11.12.20 Work on Computer Ex. 1 Consistency, Stability and Convergence Presentation Ex. 2 &
Hand out Ex. 3 + 4
7. 14.12.20-18.12.20 Submit & Present Computer Ex. 1 Finite Volume Method 1 Work on Ex. 3 + 4
Computer Ex. 2
8. 21.12.20-23.12.20 Work on Computer Ex. 2 Finite Volume Method 2 Work on Ex. 3 + 4
24.12.20-06.01.21 Holiday Holiday Holiday
07.01.21-08.01.21 Presentation Ex. 3 + 4 &
Hand out Ex. 5
9. 11.01.21-15.01.21 Submit & Present Computer Ex. 2 Numerical solution of the Navier-Stokes-Equations 1 Work on Ex. 5
Computer Ex. 3
10. 18.01.21-22.01.21 Work on Computer Ex. 3 Numerical solution of the Navier-Stokes-Equations 2 Work on Ex. 5
11. 25.01.21-29.01.21 Work on Computer Ex. 3 Numerical solution of the Navier-Stokes-Equations 3 Presentation Ex. 5 &
Hand out Ex. 6
12. 01.02.21-05.02.21 Work on Computer Ex. 3 Numerical solution of the Navier-Stokes-Equations 4 Work on Ex. 6
Hand out Ex.6
13. 08.02.21-12.02.21 Submit & Present Computer Ex. 3 Boundary Conditions Presentation Ex. 6

Dr.-Ing. Manuel Münsch | FAU | 2

 Erlangen

Outline

• PART 1: Repetition
• Boundary Layer: Blasius Equation

• PART 2: Time Discretization

• Introduction
• Single-Step Procedures:
• Explicit Euler, Implicit Euler, Leapfrog, Crank-Nicolson, Runge-Kutta
• Explicit versus Implicit
• Benchmark & Errors
• Multi-Step Procedures:
• Adams-Bashford, Adams-Moulton, Backward-Differencing

Dr.-Ing. Manuel Münsch | FAU | 3

PART 1: Repetition

Erlangen
 Erlangen

Boundary Layer Equations

(Source: Vorlesungsskript Numerische Strömungsmechanik, M. Breuer, LSTM, 2006)

Mass: Boundary conditions:

∂u ∂v
+ =0 (1) y =0:u=v =0 (3)
∂x ∂y
Momentum: y − > ∞ : u − > u∞ (4)
∂u ∂u ∂ 2u
u +v =ν 2 (2)
∂x ∂y ∂y
-> Parabolic in x, elliptic in y
Dr.-Ing. Manuel Münsch | FAU | 5
 Erlangen

Incompressible Newtonian Fluid

• Introduce stream-function ψ:
∂ψ ∂ψ
u= and v =− (5)
∂y ∂x
u
• Similarity assumption - the velocity profiles are matching if u∞
is a function
of the similarity coordinate
r
u u∞
= F (η) with η(x , y ) = y (6)
u∞ νx
• Find: Z
f (η) = F (η)d η (7)
• Blasius equation:
f f 00 + 2 f 000 = 0
(8)
with the boundary conditions:
• for η = 0: f 0 = f = 0 no-slip condition
• for η = ∞: f 0 = 1 non-viscid outer flow
Dr.-Ing. Manuel Münsch | FAU | 6
 Erlangen

Boundary Layer Equations

Ordinary Differential Equation: Solution 
  (8)
Equation   can besolved
 ! numerically
  1    schemes, Euler
with: Runge-Kutta
schemes ... (see PART 2 of this lecture)

Flow velocities:
∂ψ ∂ψ ∂η
u= = = u∞f 0(η) (9)
∂y ∂η r
∂y

   ∂ψ 1 ν u∞ 0
v =− = (η f (η) − f (η)) (10)
∂x 2 x




  


      

Dr.-Ing. Manuel Münsch | FAU | 7

PART 2: Initial value problems: Time-Stepping

Erlangen
 Erlangen

Example: Classical time dependent flow

Flow around a cylinder at Re = 200
0.02

0.015 U/U∞
1.5
0.01 1.4
1.3
1.2
0.005 1.1
1
Y [m]

0.9
0 0.8
0.7
0.6
-0.005 0.5
0.4
-0.01 0.3
0.2
0.1
-0.015

-0.02
0 0.02 0.04
X [m]

• vortical structures are shedding from the surface of the cylinder leading to
the formation of a vortex street
Dr.-Ing. Manuel Münsch | FAU | 9
 Erlangen

Introduction
Navier-Stokes Equations
For discretization purposes the conservation equations for mass and
momentum are casted into a single transport equation:

∂(ρφ) ∂(ρuj φ) ∂ ∂φ
+ = (Γ ) + Sφ (11)
∂t ∂ xj ∂ xj ∂ xj
with the following variables:

Conserv. Eq. φ Γ Sφ

Mass 1 0 0

∂ uj
Momentum ui µ − ∂∂xpi + ∂
∂ xj (µ ∂ xi ) + ρgi
∂ uj
with ∂ xj = 0:

Momentum ui µ − ∂∂xpi + ρgi

Dr.-Ing. Manuel Münsch | FAU | 10

 Erlangen

Introduction
Incompressible Navier-Stokes Equations are:
• elliptic in space (x , y , z ):
• a disturbance of the flow field at one particular point has an impact on the entire
flow field
• every location inside the flow field is influenced by all other locations inside the
flow field
• are parabolic in time (t ):
• a disturbance of the flow field has just an impact on the future development of
the flow field
• these characteristic properties are considered via Time Marching
Approaches
-> the solution is advanced in time step by step
• Time Marching Approaches are very similar to general solution strategies of
initial values problems (like Blasius equation)

Dr.-Ing. Manuel Münsch | FAU | 11

 Erlangen

Introduction
Time Marching
• Time Marching Approaches are very similar to general solution strategies
of initial values problems (like Blasius equation) which becomes more
obvious by a reformulation of the equation:

∂(ρφ) ∂(ρuj φ) ∂ ∂φ
=− + (Γ ) + Sφ = Res(φ(t ), t ) (12)
∂t ∂ xj ∂ xj ∂ xj
with residual Res(φ(t ), t ) containing all convective and diffusive fluxes

• for comparison: ordinary differential equation

d φ(t )
= f (φ(t ), t ) with φ(t0) = φ0 (13)
dt

Dr.-Ing. Manuel Münsch | FAU | 12

 Erlangen

Introduction
Time Marching
Time Marching Schemes are divided into Single-Step (or two-level) methods
and Multi-Step (or multipoint) methods and furthermore into explicit and
implicit approaches:
• Single-Step (or two-level) methods:
• just an old and a new time domain are used to predict φ(tn+1 )

φ(tn+1) = f (φ(tn+1), φ(tn )) (14)

• Multi-Step (or multipoint) methods:

• multiple time domains are used to approximate the integration in time

φ(tn+1) = f (φ(tn+1), φ(tn ), φ(tn−1), ...) (15)

Dr.-Ing. Manuel Münsch | FAU | 13

 Erlangen

Introduction
Time Marching
Time Marching Schemes are divided into Single-Step (or two-level) methods
and Multi-Step (or multipoint) methods and furthermore into explicit and
implicit approaches:
• explicit approach:
• only old time domains are used to predict the function f

φ(tn+1) = f (φ(tn ), φ(tn−1), φ(tn−2), ...) (16)

• implicit approach:
• old and new time domains are used to predict f

φ(tn+1) = f (φ(tn+1), φ(tn ), φ(tn−1), ...) (17)

Dr.-Ing. Manuel Münsch | FAU | 14

 Erlangen

Time Marching: Simple Time Stepping Schemes

Starting Point:
• ordinary differential equation (ODE) with initial condition φ(t0) = φ0
d φ(t )
= f (φ(t ), t ) with φ(t0) = φ0 (18)
dt
Task:
• estimate solution of the ODE at time t1 = t0 + ∆t basing on initial condition
φ = φ0 and time-step size ∆t
φ1 = φ(t1) = φ(t0 + ∆t ) (19)

• once the solution φ1 is known: φ1 is used as initial condition for the

succeeding time-step
φ2 = φ(t2) = φ(t1 + ∆t ) (20)

• the procedure is repeated until a solution for the desired physical time is
obtained
Dr.-Ing. Manuel Münsch | FAU | 15
 Erlangen

Time Marching: Single-Step Procedure

Derivation:
• integration from time tn to tn+1:
tn+1 tn+1 t n +1
d φ(t )
Z Z Z
n+1 n
dt = d φ(t ) = φ −φ = f (φ(t ), t )dt (21)
tn dt tn tn

• the upper equation is an exact relation if f (φ(t ), t ) is known

• this is not the case, since f (φ(t ), t ) depends on the unknown solution φ(t )
R t n +1
-> general approximation for tn f (φ(t ), t )dt required:
Z tn+1
φn+1 − φn = f (φ(t ), t )dt ≈ f (φ(τ ), τ )∆t (22)
tn

τ has to be in the interval: tn ≤ τ ≤ tn+t

• the arbitrary time
• basing on the choice of τ different solution approaches arise (next slides)

Dr.-Ing. Manuel Münsch | FAU | 16

 Erlangen

Time Marching: Single-Step Procedure

Explicit Euler Scheme
• choosing τ = tn
φn+1 = φn + f (φn , tn )∆t + T (∆t 1) (23)

• here the variable f (fluxes & source terms) is evaluated basing on the
known value φn
• the only remaining unknown is φn+1 which can be computed directly
-> explicit approach
• formally this is a forward difference in time with truncation error of first
order (Taylor Series)

Dr.-Ing. Manuel Münsch | FAU | 17

 Erlangen

Time Marching: Single-Step Procedure

Implicit Euler Scheme
• choosing τ = tn+1
φn+1 = φn + f (φn+1, tn+1)∆t + T (∆t 1) (24)

• here the variable f (fluxes & source terms) is evaluated basing on the
unknown value φn+1 at tn+1
• the only known variable is φn
• φn+1 cannot be computed directly, instead an equation system has to be
solved
-> implicit approach
• formally this is a backward difference in time with truncation error of first
order (Taylor Series)

Dr.-Ing. Manuel Münsch | FAU | 18

 Erlangen

Time Marching: Single-Step Procedure

Leapfrog Scheme
• choosing τ = tn+1/2

φn+1 = φn + f (φn+1/2, tn+1/2)∆t + T (∆t 2) (25)

• here f is evaluated basing on a intermediate time level tn+1/2 with

tn ≤ tn+1/2 ≤ tn+t
• thus the integral is evaluated basing on the midpoint rule
• since f is a function of φn+1 the variable φn+1 cannot be computed directly
-> implicit approach
• since the midpoint rule is of second order also the Leapfrog Scheme is of
second order accuracy / truncation error of second order (Taylor Series)

Dr.-Ing. Manuel Münsch | FAU | 19

 Erlangen

Time Marching: Single-Step Procedure

Crank-Nicolson Scheme
• choosing τ =?
n+1 1 n n+1
= φ + f (φ , tn ) + f (φ , tn+1) ∆t + T (∆t 2)
n

φ (26)
2

• here the integral is evaluated basing on the trapezoidal rule

• since f is a function of φn+1 the variable φn+1 cannot be computed directly
-> implicit approach
• since the trapezoidal rule is of second order also the Crank-Nicolson
Scheme is of second order accuracy / truncation error of second order
(Taylor Series)

Dr.-Ing. Manuel Münsch | FAU | 20

 Erlangen

Time Marching: Single-Step Procedure

Interpolation Scheme
• the explicit Euler-, the implicit Euler- and the Crank-Nicolson Scheme can
be casted into a single interpolation formula:
n+1 n+1
= φ + (1 − αt )f (φ , tn ) + αt f (φ , tn+1) ∆t + T (∆t 2)
n n
 
φ (27)

• depending on the value of αt within the interval [0, 1] the particular

scheme is chosen:

• αt = 0:
-> explicit Euler Scheme
• αt = 1:
-> implicit Euler Scheme
• αt = 0.5:
-> Crank-Nicolson Scheme

Dr.-Ing. Manuel Münsch | FAU | 21

 Erlangen

Time Marching: Single-Step Procedure

Additional Schemes
• in literature a lot more schemes are known, for example:
• Lax-Keller
• Lax-Wendroff
• Mac-Cormack
• implicit/explicit Runge-Kutta
• ... (see additional literature on numerical mathematics or CFD)

• among these, the explicit Runge-Kutta scheme is frequently used -

especially for cases where small time step sizes are required to resolve
physics: the prediction of turbulent flows for example

Dr.-Ing. Manuel Münsch | FAU | 22

 Erlangen

Time Marching: Runge-Kutta Schemes

Example: Classical 4-step Runge-Kutta
Idea: derive φn+1 by several sub-steps within the interval ∆t
• Step 1: predictor step basing on an explicit Euler
(1) n ∆t
φ =φ + f (φn , tn ) (28)
2
• Step 2: corrector step basing on an implicit Euler
∆t
φ(2) = φn + f (φ(1), tn+1/2) (29)
2
• Step 3: predictor step basing on the midpoint rule
(3) n ∆t (2)
φ = φ + f (φ , tn+1/2) (30)
1
• Step 4: final corrector step basing on Simpsons rule
n ∆t
h i
n+1 n (1) (2) (3)
φ =φ + f (φ , tn ) + 2f (φ , tn+1/2) + 2f (φ , tn+1/2) + f (φ , tn+1)
6
(31)
Dr.-Ing. Manuel Münsch | FAU | 23
 Erlangen

Time Marching: Runge-Kutta Schemes

Example: Classical 4-step Runge-Kutta
Idea: derive φn+1 by several sub-steps within the interval ∆t
• in general an N-step explicit Runge-Kutta Scheme offers an order of
convergence of O(∆t N ) up to N = 4
• for N ≥ 4 the order of convergence still increases but not linearly with N
anymore
• the classical Runge-Kutta schem has high storage requirements since all
sub-step results have to be solved
-> drawback for solving large problems
-> to solve this Low-Storage Runge-Kutta schemes have been developed

Dr.-Ing. Manuel Münsch | FAU | 24

 Erlangen

Time Marching: Runge-Kutta Schemes

Example: Low-Storage Runge-Kutta (3 sub-steps)
• developed to reduce storage amount of sub-steps

1
φ(1) = φ(n) + α1∆t f (φn , tn ) with α1 =
3
1
φ(2) (n) 1
= φ + α2∆t f (φ , t1) with α2 = (32)
2
φ(3) = φ(n) + α3∆t f (φ2, t2) with α3 = 1
φn+1 = φ(3)

Dr.-Ing. Manuel Münsch | FAU | 25

 Erlangen

Time Marching: Runge-Kutta Schemes

Example: Low-Storage Runge-Kutta (k sub-steps)
• general formulation for k sub-steps:

φ(1) = φ(n) + α1∆t f (φn , tn )

φ(2) = φ(n) + α2∆t f (φ1, t1)
...
(33)
φ(k −1) = φ(n) + α(k −1)∆t f (φ(k −2), tk −2)
φ(k ) = φ(n) + α(k )∆t f (φ(k −1), tk −1)
φn+1 = φ(k )
• for the Low-Storage formulation the order of convergence reduces to
O(∆t 2) for N ≥ 4 when function f is non-linear
• the number of steps k increases the stability of the scheme
• order of convergence and stability also depend on the factors α(k )

Dr.-Ing. Manuel Münsch | FAU | 26

 Erlangen

Time Marching: Runge-Kutta Schemes

Example: Low-Storage Runge-Kutta (k sub-steps)
• these factors can be adjusted accordingly (see literature)
• here α(k ) is evaluated basing on:

1
αk = with k = 1, 2, ..., N (34)
N −k +1

• in general the usage of explicit Runge-Kutta Schemes is restricted to small

time step sizes due to stability reasons (link to Lecture on Stability)
• these schemes are mostly used for the prediction of turbulent flows where
small time step sizes are required to resolve turbulent fluctuations (link to
NMTFD2)

Dr.-Ing. Manuel Münsch | FAU | 27

 Erlangen

Time Marching: Single-Step Procedure

Example Benchmark
To validate the performance and implementation of any numerical scheme a
comparison with a known or analytical solution of a benchmark problem is
highly recommended.
• Ordinary Differential Equation (ODE):
du
= f (u , t ) (35)
dt

• together with an initial condition it is an Initial Value Problem (IVP):

du
= f (u , t ) with u (0) = uo (36)
dt
• test function: f (u , t ) = λu with λ = const . and u (0) = 1 has the analytical
solution
u (t ) = eλt (37)

Dr.-Ing. Manuel Münsch | FAU | 28

 Erlangen

Time Marching: Single-Step Procedure

Example Benchmark
• Explicit Euler method: un+1 = un + ∆t fn

• applying Explicit Euler to solve the problem: f (u , t ) = λu with λ = −1 and

u (0) = 1

• leads to the following implementation:

dt=0.1
u(1)=1.0
l=-1.0
do i=1,N
u(i+1) = u(i) + dt*l*u(i)
end do

Dr.-Ing. Manuel Münsch | FAU | 29

 Erlangen

Time Marching: Single-Step Procedure

Example Benchmark
• Implicit Euler method: un+1 = un + ∆t fn+1

• at each time-step we must now solve the equation

un+1 = un + ∆t f (un+1, tn+1) (38)

• for the benchmark problem: f (u , t ) = λu with λ = −1 and u (0) = 1

• this leads to:

un
un+1 = un + ∆t λun+1 −→ un+1 = (39)
1 − λ∆t

Dr.-Ing. Manuel Münsch | FAU | 30

 Erlangen

Time Marching: Single-Step Procedure

Example Benchmark
• Implementation:
dt=0.1
u(1)=1.0
l=-1.0
do i=1,N
u(i+1) = u(i) / (1.d0 - l*dt)
end do
• Remark: due to the relatively simple benchmark problem the Implicit Euler
is as easy to solve as the Explicit Euler
-> in general, especially for nonlinear equations this is not the case!!!
• for example:

f (u , t ) = cos(u ) −→ un+1 = un + ∆t cos(un+1) (40)

• this equation is nonlinear and must be solved numerically at each step
with -> Newton’s method
Dr.-Ing. Manuel Münsch | FAU | 31
 Erlangen

Time Marching: Single-Step Procedure

Solution Explicit Euler (I): Comparison of result with Explicit Euler method

Exact Solution
Explicit Euler dt = 2.1429
6 Explitcit Euler dt = 1.875
Explitcit Euler dt = 1.5
Explitcit Euler dt = 0.1

2
Solution u

-2

-4

-6

0 5 10 15 20 25 30
Time

Dr.-Ing. Manuel Münsch | FAU | 32

 Erlangen

Time Marching: Single-Step Procedure

Solution Explicit Euler (II): Discussion
• the Explicit Euler Method shows unstable behavior for larger time-step
sizes
• The Explicit Euler method applied to the equation
du
= f (u , t ) with u (0) = uo (41)
dt
is stable for:
2
∆t < here: ∆t < 2s (42)
|λ|
• in general, the Explicit Euler method will converge only if the time-step ∆t
is small enough
• similar restrictions apply to all explicit methods

Dr.-Ing. Manuel Münsch | FAU | 33

 Erlangen

Time Marching: Single-Step Procedure

Solution Implicit Euler: Comparison of result with Explicit Euler method
1
Exact Solution
Implicit Euler dt = 2.1429
Implicit Euler dt = 1.875
Implicit Euler dt = 1.5
Implicit Euler dt = 0.1
0.8

0.6
Solution u

0.4

0.2

0 5 10 15 20 25 30
Time

Dr.-Ing. Manuel Münsch | FAU | 34

 Erlangen

Time Marching: Single-Step Procedure

Solution Crank-Nicolson: Comparison of result with Crank-Nicolson method
1
Exact Solution
Crank-Nicolson dt = 2.1429
Crank-Nicolson dt = 1.875
Crank-Nicolson dt = 1.5
Crank-Nicolson dt = 0.1
0.8

0.6
Solution u

0.4

0.2

0 5 10 15 20 25 30
Time

Dr.-Ing. Manuel Münsch | FAU | 35

 Erlangen

Time Marching: Single-Step Procedure

Solution Runge-Kutta (4-step): Comparison of result with RUK4 method
1
Exact Solution
RUK4 dt = 2.1429
RUK4 dt = 1.875
RUK4 dt = 1.5
RUK4 dt = 0.1
0.8

0.6
Solution u

0.4

0.2

0 5 10 15 20 25 30
Time

Dr.-Ing. Manuel Münsch | FAU | 36

 Erlangen

Time Marching: Single-Step Procedure

Errors (I):
• the performance of an implemented algorithm is evaluated by considering
the mean absolute error for example
• the mean absolute error ε is estimated by considering the normalized
difference between numerical and analytical solution divided by the
number of solutions N basing on an identical physical time interval:
N
1X n
ε= |(u − uanaly )| (43)
N n=1
• the gradient p of the error curves is equal to the order of accuracy of the
underlying algorithm (logarithmic display)
log(ε(∆t1)) − log(ε(∆t2))
p= (44)
log( ∆1t1 ) − log( ∆1t2 )
-> use this strategy to check the quality of your coding
Dr.-Ing. Manuel Münsch | FAU | 37
 Erlangen

Time Marching: Single-Step Procedure

Errors (II): Mean Absolute Error versus Time

Dr.-Ing. Manuel Münsch | FAU | 38

 Erlangen

Time Marching: Explicit versus Implicit

Explicit Implicit
• the quantity f can be directly • the quantity f cannot be computed
computed without any iterations directly: iterative approach
________
• solution of a linear equation system • solution of a non-linear
linear equation system
not required required
• easy to code • complex to code
• straightforward implementation on • complex implementation on HPC
HPC (vectorization & parallelisation) (vectorization & parallelisation)
• low numerical effort per time-step in • significant higher numerical effort
terms of CPU-time and storage per time-step in terms of CPU-time
requirements and storage requirements
• strong restriction on the time step • in general no restriction on the time
size: only stable for ∆t ≤ ∆tstable step size: unconditional stable
(linear model equation)
• larger time step sizes applicable:
potential compensation of higher
Dr.-Ing. Manuel Münsch | FAU | numerical effort per time step 39
 Erlangen

Time Marching: Multi-Step Procedure

Introduction
• within the Multi-Step Procedure multiple time domains are used
• analogous to the spacial discretization in the FD or FV context a
polynomial function is chosen to approximate f (φ(t ), t )
• according to the order of the polynomial function (the number of time
domains) a variety of formulation is obtained
- the Multi-Step Procedure always has to be started with a Single-Step
approach until sufficient time domains are estimated
- the data of several time domains has to be stored -> high storage
requirements

Dr.-Ing. Manuel Münsch | FAU | 40

 Erlangen

Time Marching: Multi-Step Procedure

Explicit: Adams-Bashforth-Scheme
Within explicit Multi-Step Procedures the right hand side of equation (22) is
approximated with a polynomial approach basing on known time domains
(with ∆t = const .):

• Polynomial of zeroth order: Adams-Bashforth-Scheme (1. order)

φn+1 = φn + ∆t f (φn , tn ) (45)

-> corresponds to Explicit Euler, thus still a Single-Step procedure

• Polynomial of first order: Adams-Bashforth-Scheme (2. order)

n+1 n ∆t  n n−1

φ =φ + 3f (φ , tn ) − f (φ , tn−1) (46)
2

-> most used Adams-Bashforth formula

Dr.-Ing. Manuel Münsch | FAU | 41

 Erlangen

Time Marching: Multi-Step Procedure

Explicit: Adams-Bashforth-Scheme
• Polynomial of second order: Adams-Bashforth-Scheme (3. order)

n+1 n ∆t  n n−1 n−2


φ =φ + 23f (φ , tn ) − 16f (φ , tn−1) + 5f (φ , tn−2) (47)
12

• Polynomial of third order: Adams-Bashforth-Scheme (4. order)

∆t
φn+1 = φn + [55f (φn , tn ) − 59f (φn−1, tn−1)
24
+37f (φn−2, tn−2) − 9f (φn−3, tn−3)] (48)

Dr.-Ing. Manuel Münsch | FAU | 42

 Erlangen

Time Marching: Multi-Step Procedure

Implicit: Adams-Moulton-Scheme
Within implicit Multi-Step Procedures the right hand side of equation (22) is
approximated with a polynomial approach basing on new and old time
domains (with ∆t = const .):

• Polynomial of zeroth order: Adams-Moulton-Scheme (1. order)

φn+1 = φn + ∆t f (φn+1, tn+1) (49)

-> corresponds to Implicit Euler, thus still a Single-Step procedure

• Polynomial of first order: Adams-Moulton-Scheme (2. order)

n+1 n ∆t  n+1 n

φ =φ + f (φ , tn+1) + f (φ , tn ) (50)
2

-> corresponds to the Crank-Nicolson approach, thus still a Single-Step procedure

Dr.-Ing. Manuel Münsch | FAU | 43

 Erlangen

Time Marching: Multi-Step Procedure

Implicit: Adams-Moulton-Scheme
• Polynomial of second order: Adams-Moulton-Scheme (3. order)

n+1 n ∆t  n+1 n n−1


φ =φ + 5f (φ , tn+1) + 8f (φ , tn ) − f (φ , tn−1) (51)
12

• Polynomial of third order: Adams-Moulton-Scheme (4. order)

∆t
φn+1 = φn + [9f (φn+1, tn+1) + 19f (φn , tn )
24
−5f (φn−1, tn−1) + f (φn−2, tn−2)] (52)

Dr.-Ing. Manuel Münsch | FAU | 44

 Erlangen

Time Marching: Multi-Step Procedure

Implicit: Backward-Differencing-Formula
Further implicit Multi-Step Procedures can be derived by expressing the right
hand side of equation (22) via backward differences of desired order of
accuracy (with ∆t = const .):

• 1. order approach
φn+1 − φn
= f (φn+1, tn+1) (53)
∆t
-> corresponds to Implicit Euler of first order, thus still a Single-Step procedure

• 2. order approach: Three-Point-Backward Scheme

3φn+1 − 4φn + φn−1

= f (φn+1, tn+1) (54)
2∆t

-> commonly used implicit Multi-Step Procedure

-> compared to Implicit Euler one additional time domain to be solved but
improvement of stability
Dr.-Ing. Manuel Münsch | FAU | 45
 Erlangen

Time Marching: Multi-Step Procedure

Implicit: Backward-Differencing-Formula
• 3. order approach:

11φn+1 − 18φn + 9φn−1 − 2φn−2

= f (φn+1, tn+1) (55)
6∆t

• 4. order approach:

25φn+1 − 48φn + 36φn−1 − 16φn−2 + 3φn−3

= f (φn+1, tn+1) (56)
12∆t

Dr.-Ing. Manuel Münsch | FAU | 46

 Erlangen

Time Marching
Conclusion
• the choice of an optimal time stepping algorithm is strongly problem
dependent
• the selection depends on
• the required accuracy of the solution for a particular problem
• the stability of the algorithm
• the numerical effort regarding CPU-time and storage requirements
• for unsteady problems with strong variations of the flow field 2nd order
approaches should be chosen
• for problems requiring small time-step sizes to resolve the physics explicit
approaches may be an option
• for stiff problems implicit methods (Crank-Nicolson, BDF, Adams-Moulton)
are in general the best choice

Dr.-Ing. Manuel Münsch | FAU | 47