Professional Documents
Culture Documents
simulation convergence
March 2014
1
Presentation Plan
Introduction
High Performance Computing
Mesh Quality
ANSYS Solvers
Numerical Schemes
High order terms relaxation
Reduced Rank Extrapolation
Pseudo Transient Method
PBNS Solver settings for external aerodynamics
Special settings for thermal applications
Convergence Acceleration for stretched mesh
Conclusion
2
Introduction
Introduction
ANSYS Solver is committed to deliver the best-in-class solvers
These solvers rely on 3 technologies :
Hardware
Numerical schemes
Physical models
% of GPU
14000
12000
18
16
14
10000
12
8000
10
6000
4000
2000
0
2009
2010
2011
2012
2013
number
cores/processor
0
2009
2010
2011
2012
2013
0
2009
2010
2011
2012
2013
TOP500.Org
Fluent
2048
1536
700
15.0
600
14.5
Rating
800
ideal
1024
500
400
300
512
R15.0
200
Ideal
100
0
0
512
1024
Number of Cores
1536
2048
0
0
Number of Cores
Rating is defined as the number of benchmarks that can be run on a given machine (in
sequence) in a 24 hour period. It is computed by dividing the number of seconds in a day by the
number of seconds required to run the benchmark. A higher rating means faster performance.
7
Fluent
25000
12.0.19
Performance Rating
20000
13.0.1
15000
10000
5000
0
0
64
128
192
256
320
Num Cores
384
448
512
P6
P2
P7
P1
P8
P3
P5
P1
P3
P2
P7
P5
P6
P4
P8
Compute Node 1
Compute Node 2
Rating
5000
4000
MPI
Hybrid
2Domain
3000
2000
1000
0
16
10
32
64
128
256
384
512
Fluent
1.1 million
surface clusters
0.4 million
surface clusters
11
R13
R14
14.5.0
15.0.0
6000
5000
Time in seconds
4000
3000
2000
1000
0
1024
2048
4096
6144
8192
9216 10240
13
architecture
Support for Intel ManyIntegrated-Core
architecture
13.0.0
14.0.0
15.0.0
2000
1500
1000
500
0
0
14
2048
4096
6144
NumCores
8192
10240
12288
Mesh quality
15
Meshing strategy
Accuracy
Efficiency
16
Simplicity
Meshing guidelines
Desired mesh quality
What is the maximum
skewness and aspect
ratio you can tolerate?
Time available
Faster Tet-dominant mesh
vs crafted Hex/hybrid
mesh with lower cell
count
Mesh quality
Good
Cell not too
distorted
Cell not too
stretched
Smooth
Cells
transition
18
Not Good
19
Boundary layers:
Accuracy comparison
Hex vs Tet mesh
Quad/Hex aligned with the flow are more accurate than Tri
Without dominant flow direction Quad & Tri equivalent
Contours of axial
velocity magnitude
U=0.1
U=1.0
U = V = 1.0 , T = 0
20
Tri mesh
U = V = 1.0, T = 1
qu
ad
U = V = 1.0,
T=1
Hex mesh
t
r
i
U = V = 1.0 , T = 0
Contours of temperature
for inviscid flow
ANSYS Solvers
21
ANSYS Solvers
All ANSYS solver have an all-Mach formulation.
Each of these solvers may be more effective on specific
problems
Choosing the right solver for the right application is the
first step to a fast simulation convergence
ANSYS solvers can be separated in two technologies
depending on which form the continuity equation is
solved :
Pressure based Navier-Stokes solvers (PBNS)
Density based Navier-Stokes solvers (DBNS)
Traditionnaly, PBNS is more suited for incompressible
flows and DBNS for compressible
22
ANSYS Solvers
Each of these technologies are also separated depending
on how equations are coupled together
Pressure Based
Fluent
Implicit Segregated
Fluent
CFX
All equations
are solved in a segrageted way
Fluent
Implicit Coupled
Fluent
Density Based
23
Implicit Coupled
Continuity and Momentum
are solved in coupled way
Explicit Coupled
All equatiions
are solved in a coupled way
ANSYS Solvers
Each of these technologies are also separated depending
on how equations are coupled together
Pressure Based
Implicit Segregated
Fluent
Fluent
Fluent
CFX
Fluent
Density Based
2011 ANSYS, Inc.
Mach < 2
Most effective solver
Explicit Coupled
Mach > 2
Unsteady flows
Mach > 2
Most effective solver
24
Implicit Coupled
Pressure-based Solvers
(FLUENT & CFX)
Pressure-Based
25
Pressure-Based
(segregated)
Pressure-Based
(coupled)
Solve U-Momentum
Solve V-Momentum
Solve W-Momentum
Solve Mass
& Momentum
Solve Mass
Continuity;
Update Velocity
Solve Energy
Solve Species
Solve Turbulence Equation(s)
Solve Other Transport Equations as required
Density-based Solvers
(FLUENT)
Density-Based
Density-Based
(coupled implicit)
Density-Based
(coupled explicit)
Solve Mass,
Momentum,
Energy,
Species
Solve Mass,
Momentum,
Energy,
Species
26
Numerical Schemes
27
28
29
With
No Impact on results
Without
30
31
Without
32
RRE
Reduced Rank Extrapolation should be
used to accelerate convergence of the
slowly converging cases
Available as a beta feature since 14.5
It is a vector extrapolation method using
the previous convergence steps
Inputs are :
The number of steps to use for extrapolation
(ie Subspace size)
The frequency of use of RRE
33
34
35
with
1
1
CFL
n 1
p
pn
a
p
n 1
p
ai
n 1
i
Sp
with
length _ scale
velocity _ scale
Pseudo Timestep
The timestep is used to move the solution towards the final
answer
Relaxation of the equation non-linearities
Transient evolution of the flow from the initial guess to the
steady-state conditions
Converged solution is independent of the timestep used
Initial Guess
50 iterations
100 iterations
150 iterations
Final Solution
37
CFX
Fluent
Auto Timescale,
Automatic
Physical Timescale
User Defined
38
, ,
40
max( , , )
0.3
max( , )
0.3
0.1
0.3
max( , , , )
where =
is the diffusity
41
User Timescale
42
CFX
Fluent
43
Transient effect
Sometimes simulations which are run in steady state mode
will not converge even with good mesh quality and well
selected timestep.
If a steady state run shows oscillatory behavior of the
residual plots, a good test is to reduce or increase the
timestep by known factors.
If the period of oscillation of the residual plot changes by
changing the timestep, then the phenomenon is most likely
a numerical effect.
If the period stays the same, then it is probably a transient
effect.
In Fluent, switching back to a steady formulation might
reduce this problem
44
750
75
2300
1350
1200
100
500
250
220
400
50
110
CFX
Fluent
48
Case
PBCS
Pseudo Transient off
PBCS
Pseudo Transient on
PBCS
Pseudo Transient on
But off for Species & Energy
Iter.
244
125
66
Initialization
49
50
Standard
Hybrid
Full
MultiGrid
Previous
Calculation
Standard initialization
Initialization is done based on specific value
Boundary value can both be specified or calculated
from boundary conditions
Fluent
CFX
51
Patch
Patch values for individual variables in certain regions
Free jet flows (high velocity for jet)
Combustion problems (high temperature
region to initialize reaction)
Cell registers (created by marking the
cells in the Adaption panel) can be used
for patching values into various regions
of the domain.
Multiphase flows
(patch different phase
volume fractions in
one or more regions)
52
Fluent
POUTLET(Auxiliary Out)
P = 0.0
POUTLET(Primary Out)
P = 0.0
MFINLET(Primary In)
MFR = 1.14; T0 = 322.04 K
Case Setup :
54
Initialization Fields
55
Hybrid Init:
Iterations = 102
URF
Mom 0.7, Press 0.3, Den 1.0
Energy 1.0
FMG tips
Make sure before using FMG you have performed proper standard
or hybrid initialization (see previous slides)
Use the FMG verbosity, so you see convergence behavior
Examine the guessed solution from FMG before you proceed with
normal iterations.
In general you want to perform more cycles on coarse grids than
fine grids
Using too many grid levels can be problematic in some flow
topology:
Coarsest level may create single cells in thin passages leading to
break up in solution.
FMG Example
Numerical solution initialized from the free-stream flowfield
Full multigrid (FMG) initialization applied to obtain the initial solution
60
61
Hyb init
FMG init
V max = 55 m/s
62
V max = 213.6m/s
Interpo init
Velocity field is
predicted quite well
in both cases
Comparison
Initialization Number of
iteration to
convergence
63
Standard
600
Hybrid
500
FMG
250
Previous Cal.
150
All theses case have run on the same type of machines and
on the same number of procs.
The convergence process, looking at the residual, is
identical for the Std and Hyb init. While it is different for
the FMG & Ip.
Convergence is reached quicker with FMG and Ip in less
than 250 iter
while 500 iter is needed for the Hyb & Std.
2011 ANSYS, Inc.
April 17, 2014
Gig
sec
Sec
Comparison
65
Introduction
Switching the pseudo-transient solver option can lead to
slower convergence on large cases
The pseudo transient term introduction can trigger the solver
sensibility to instabilities
Aerodynamics variables like drag or lift were observed to
oscillate in some cases using SST k-w
An alternative approach is to use the classic coupled PBNS
solver with F-Cycle for Turbulence, in a two or three steps
approach and play on CFL and under relaxation factors (high
for Turbulence equations)
Observed Outcome: faster convergence and reduced
oscillation of SST k-w
66
+ +
Explicit: = +
The CFL number input inside fluent is a way to control all implicit
under-relaxation in the resolved equation at the same time
=
1
=
1 +
67
Step 1
Step 2
Step 3
CFL
50
200
100
Pressure
explicit under relaxation
0.25
0.5
0.25
Pressure
explicit under relaxation
0.25
0.5
0.25
Turbulence
Implicit under relaxation
0.8
0.95
0.95
% of iterations
60
35
2 steps strategy
68
Step 1
Step 2
CFL
100
200
Pressure
explicit under relaxation
0.4
0.4
Pressure
explicit under relaxation
0.4
0.4
Turbulence
Implicit under relaxation
0.8
0.95
% of iterations
20
80
3 steps
Chord = 299 mm
Span width = 150 mm
Angle of attack alpha = 13.86 deg.
Re ~ 1,000,000
About 3,000,000 cells, hybrid mesh (prism+tets)
SST kw
2 steps
Realizable k-e
SST k-w
SST k-w
with 3 step strategy
70
Solver settings
for thermal applications
71
Double-Precision Solver
The double-precision solver is designed to minimize truncation error and
thus improve the overall heat balance.
Fluent
CFX
Double precision doubles the memory need but only increases by 10% the
calculation time
72
CFX
Fluent
OR
73
Energy residuals
74
F-Cycle
Energy residuals
Under-Relaxation of Energy
Under-relaxation of energy equation can be implicit and explicit :
Implicit:
+ +
Explicit: = +
Generally if implicit URF is reduced to even slightly (to 0.99), it will
take many iterations to converge.
Instead, explicit URF can be reduced to as low as 0.1 and still obtain
convergence in reasonable number of iterations.
CFX
Fluent
75
Explicit Under-Relaxation
15 million underhood case
explicit URF=0.5
76
Gradient Schemes
Gradients of solution variables are required in order to evaluate diffusive fluxes,
velocity derivatives, and for higher-order discretization schemes.
Fluent
Trilinear (default)
Linear-Linear
77
5mm
Symmetry
5.103 .9000
=
+50
0.1
= 500
=
Fluent GGCB
0 K difference
Fluent LSCB
0 K difference
15 K difference
Fluent GGNB
4 K difference
Secondary Gradients
What is a secondary gradient?
Secondary gradient is introduced when a cell is skewed.
Disable this secondary gradient can help on convergence
for poor quality mesh
q k T n
T T
k w c f T
h
Tc
r
Tw
Tc
Tw
80
Secondary gradient
Secondary Gradients
How to disable secondary gradients
Disabling secondary gradient only adjacent to walls (Alternative wall
formulation)
/solve set expert , yes , , ,
81
Secondary Gradients
Do we loose accuracy by ignoring secondary gradient?
Typically, highly skewed cells are located in areas of less importance
(unresolved gaps, corners, etc.)
Thus, accuracy is not compromised if proper meshing guidelines are
followed.
Default
82
Secondary Gradients
50K
T?
Q=9,000 W/m^2
T(analytic)=9000*.005/.1+50=500K
Error = 42 %
83
LSCB
Error = 0 %
Speed-up conduction
convergence for transient
Thermal Applications
84
Biot number
85
Fourier number
Speedup Factor
A speedup factor f can be introduced so that : =
86
real time
89
Convergence Acceleration
For Stretched Meshes (CASM)
Accelerate the convergence of the
DBNS implicit on highly-stretched
meshes.
Convergence can be between 2 to 10
times faster than without using CASM.
Use CFL value multiplied by a factor
proportional to cell aspect ratio.
Cell stretched perpendicular to flow
skipped
Steady-State solution
Can be used with Solution-Steering
but Manual schedule adjustment is
required.
90
lmax
lmin
CFL
lmin
A
f
AR CFL
lmax
CASM
CASM option can be found in the
solver methods description.
When CASM is in use you typically
do not need to run the solver at very
high CFL value. Range between 2 to
50 is sufficient for most flow cases.
FMG initialization should be used
before solving flow with CASM
91
Mesh properties
3.5 million hex cell mesh
Max AR = 2.6e06
92
93
200
2000
Conclusion
94
Conclusions
ANSYS CFD is working on the basis of five solvers
Each solver should be used knowing it most favourable domain
of application
Multiple now techniques exists to increase solver efficiency :
95