Professional Documents
Culture Documents
Phase 1
Detection of Butterfly Effect in OpenFOAM
In previous runs done by SL-RASCH with PAM-FLOW (FEM CFD solver) to extract
moments and forces on structures, it was observed that flow development process
and instantaneous results would change even with seemingly innocuous
modifications in the setup of the simulations. Items like number of CPUs and
partitioning had a significant impact onto these instantaneous results.
If there was no change at all in the simulation, the results did not deviate. However,
changing the number of CPUs (and hence partitioning) had a significant impact
onto results. Round-off errors were highlighted as the probable reason for such
behaviour.
This project was performed to check if the same phenomenon appeared in
OpenFOAM (FVM CFD solver).
4m cell size
2m cell
size
2m cell
size
0.5m 0.25m
cell size cell size
Copyright © ESI Group, 2012. All rights reserved. 7
Model Description:
X Section
4m cell
size
2m cell
size
60m
0.25m
cell size
94m
Copyright © ESI Group, 2012. All rights reserved. 8
Model Description:
Inlet Profile
With the model as described, simulation was initially performed for 1800s (30 min) as
follows:
• LES with Smagorinsky Subgrid Scale model.
• Flow field initialised from a static flow (0 m/s in all directions).
• Simulation was done with 8, 32 and 64 CPUs. Partitioning was similar in all cases:
simple partitioning method (coefficients 4-2-1, 16-2-1 and 32-2-1 in X-Y-Z). Set-up
files were the same for all 3 cases.
• Time step sizes of 5ms (5e-03s), which produced maximum courant numbers of
1.5-2.0 along time.
• Following results were extracted from the simulation:
• Forces and moments on the umbrella were extracted for each time step.
• Velocity and static pressure values on monitoring points in the domain were
extracted for each time step.
• Flow field (p and U) for every 5 seconds.
Copyright © ESI Group, 2012. All rights reserved. 10
Model Description:
Set-up and Simulation
Point #3
As observed in the cases done with PAM-FLOW, the first 12 seconds of simulation
showed a very similar set of results for the cases with different CPUs.
However, after this time, the calculated instantaneous force progressed along
completely different curves.
Similarly with the moments, the first 10-13 seconds of simulation showed a similar set
of results for the cases with different CPUs. After these, the calculated instantaneous
moments progressed along completely different curves.
For the first 10 seconds of simulation, the flow is similar for all cases, regardless of
the number of CPUs used.
There are some small differences around the umbrella, but the rest of the domain
looks alike.
5 seconds 10 seconds
25 s 30 s
Copyright © ESI Group, 2012. All rights reserved. 15
Results:
Velocity at points
Last ~17min
of averaging
Last ~18min
of averaging
Last ~10min
of averaging
Last ~17min
of averaging
Last ~18min
of averaging
(Isolines generated
every 50Pa)
Distribution of pressure coefficients on the umbrella are very similar regardless of the
number of CPUs used in the calculation (view from rear).
Distribution of pressure coefficients on the umbrella are very similar regardless of the
number of CPUs used in the calculation (view from front).
Distribution of pressure coefficients on the umbrella are very similar regardless of the
number of CPUs used in the calculation (view from top).
Forces and moments time-averaged over the last 1000 seconds for the different
cases were as follows:
# CPUs Fx Fy Fz
8 27838 390 -180251
32 30410 -189 -181709 Values in N
64 27885 392 -180201
# CPUs Mx My Mz
8 -15596 122808 526
32 14494 205487 -1031 Values in N*m
64 -13944 145562 1248
In terms of forces, time averaged values are very similar. However, for the moments
there is a strong variation depending on the number of cores used.
Since the average Cp values on the umbrella do not have any core dependency (as
observed in previous slides), moments should be similar in all 3 cases.
Investigation about this is shown in following slides.
Time refers to length of averaging in the 1800s of simulation time: 1000s means that only the last 1000s were included in the
averaging (from 800s to 1800s).
As the averaging time increases, the time-averaged values get closer and the relative
difference in the forces as defined by Max Min reduce except for Fy, which is close to
Max
0N.
Concerning moments, there are important differences even with averaging of 1500s,
especially for My.
Copyright © ESI Group, 2012. All rights reserved. 26
Addendum A
Investigation on Average Moment Figures
In order to check the impact of the amplitude of oscillations of the forces and
moments, RMS values of forces and moments were calculated, as follows:
Time refers to length of averaging in the 1800s of simulation time: 1000s means that only the last 1000s were included in the
averaging (from 800s to 1800s).
RMS forces and moments showed a much more consistent behavior regardless of
the number of cores used in the simulation, with maximum deviations of around 5% in
Fx and 3% in My.
Actually, since the model was symmetric around Y axis, time-averaged value of Fy,
Mx and of Mz should be very close to 0Nm.
When comparing the time-averaged values to the RMS values, the actual
significance of the time-averaged values is put in the right place:
TimeAveraged
RMS
TimeAveraged
RMS
Time refers to length of averaging in the 1800s of simulation time: 1000s means that only the last 1000s were included in the
averaging (from 800s to 1800s).
The values of forces and moments do get closer as the averaging time increases. But
the significance of the time-averaged moments is still around 20% of RMS value in
the best case, and are below 2-3% in the cases of Fy, Mx and Mz (virtually 0).
With the increase in averaging time, the RMS values are (again) consistent regardless
of the number of CPUs used. All forces and moments RMS values are within around
3% (from max).
The time-averaged values of Mx and Mz are very close to 0Nm (below 3% of the
RMS value). However, the value of My (TimeAveraged/RMS) does not converge to a
uniform value and has some variability, as shown below:
Increasing the averaging time only makes the results slightly more uniform, but even
after 3000s there are important differences between cases.
As fluctuations of forces and moments are very important, sample standard deviation
was calculated for the different cases:
Time-
averaged
values
Sample
Standard
Deviation
Comparing the figures, the standard deviations are more consistent than the time-
averaged values, and they have much higher values than the time-averaged ones
(except Fz).
Actual importance of the time-averaged values should be relativized given the strong
variability along time.
Copyright © ESI Group, 2012. All rights reserved. 32
Addendum B
Investigation on OpenFOAM solver
tolerances
Simulations with 32 and 64 CPUs were re-run for 1800s with tighter solver
tolerances in fvSolutions file, to check if the variability still appeared:
solvers solvers
{ {
p { p {
[…] […]
relTol 0.01; relTol 0.005;
[…] […]
} }
pFinal { pFinal {
[…] […]
tolerance 1.0e-06; tolerance 1.0e-07;
[…] […]
} }
U { U {
[…] […]
tolerance 1e-07; tolerance 1e-08;
[…] […]
} }
Ufinal { Ufinal {
[…] […]
tolerance 1e-07; tolerance 1e-08;
[…] […]
} }
Results from the new runs showed that the variability does not reduce when
tightening up the solver tolerances.