Note: This discussion is about an older version of the COMSOL Multiphysics® software. The information provided may be out of date.
Discussion Closed This discussion was created more than 6 months ago and has been closed. To start a new discussion with a link back to this one, click here.
Runtime of plasma simulation
Posted Mar 29, 2013, 5:44 a.m. EDT Plasma Physics Version 4.3, Version 4.3b 3 Replies
Please login with a confirmed email address before reporting spam
Dear Sir,
We have some test results as stated below after so many attempts on reducing runtime of plasma module for simulating a chamber with CCP type.
a. We tested several cases (such as argon_gec_icp and our own cases) with different number of cores. However, the parallel performance is not as good as expect, which is less than 40 % or even lower in some cases.
b. We tested the cases with both direct and iterative solvers with different preconditioners. The difference is around 30% in the cases we tested.
c. For a plasma simulation with many species (>20), we tried to solve the solution segregatedly (with several different combination), reducing the runtime dramatically (more than 10 times speedup in some cases). But it seems not good enough to resolve the difficulty to run the simulation within the “acceptable” runtime. Because of the physical constraint, we can only run the simulation with max. time step size as 10^(-10) (s) though COMSOL makes use of the fully coupled scheme which is numerically stable for larger time step size. Furthermore, it requires several thousand of RF cycles for the simulation to reach the quasi-steady solution. In other words, it may take several days up to weeks to run a simulation, which is not acceptable and practical.
The enclosure is an example of helium discharge with 25 species and 101 reactions. The element number is 110 x 50 which is not a big problem if compared with chamber simulation. If we fix the size of time step to 7x10^(-10), and run the simulation for one cycle (study 2, from 100th to 101st ). It takes about 3000 ~ 4000 (s) with fully coupled scheme, and about 400 (s) with segregated scheme. As stated previously, it will still take “weeks” to run the simulation.
We wonder whether we did anything wrong (We hope we really did). Is there any alternative (solver, segregated scheme, preconditioner, etc.) we can do to tune the example case to reach its optimized conditions? The target is to reduce the runtime to 30 second around for one cycle. Is it possible? Thanks.
KM
We have some test results as stated below after so many attempts on reducing runtime of plasma module for simulating a chamber with CCP type.
a. We tested several cases (such as argon_gec_icp and our own cases) with different number of cores. However, the parallel performance is not as good as expect, which is less than 40 % or even lower in some cases.
b. We tested the cases with both direct and iterative solvers with different preconditioners. The difference is around 30% in the cases we tested.
c. For a plasma simulation with many species (>20), we tried to solve the solution segregatedly (with several different combination), reducing the runtime dramatically (more than 10 times speedup in some cases). But it seems not good enough to resolve the difficulty to run the simulation within the “acceptable” runtime. Because of the physical constraint, we can only run the simulation with max. time step size as 10^(-10) (s) though COMSOL makes use of the fully coupled scheme which is numerically stable for larger time step size. Furthermore, it requires several thousand of RF cycles for the simulation to reach the quasi-steady solution. In other words, it may take several days up to weeks to run a simulation, which is not acceptable and practical.
The enclosure is an example of helium discharge with 25 species and 101 reactions. The element number is 110 x 50 which is not a big problem if compared with chamber simulation. If we fix the size of time step to 7x10^(-10), and run the simulation for one cycle (study 2, from 100th to 101st ). It takes about 3000 ~ 4000 (s) with fully coupled scheme, and about 400 (s) with segregated scheme. As stated previously, it will still take “weeks” to run the simulation.
We wonder whether we did anything wrong (We hope we really did). Is there any alternative (solver, segregated scheme, preconditioner, etc.) we can do to tune the example case to reach its optimized conditions? The target is to reduce the runtime to 30 second around for one cycle. Is it possible? Thanks.
KM
Attachments:
3 Replies Last Post May 29, 2013, 9:57 p.m. EDT
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
Yes, good point..
sometimes I am lucky and I can run time step of 10E-7s, but most of the time it oscillates at 10E-8 or 10E-9, in bad cases it keeps on going at 10E-16s.. which of course never comes to an end.
I am happy when I can get up to 0,1 ms of plasma.. which is not always steady-state..
it is only Argon, and segregated (Field / Species) rather big 60000 mesh units..
but I still cannot find out why some models (parameters) run faster than the other..
to me it is very sensitive to the meshing.. but I do not know how I can evaluate this. ( not really dependant on the total amount of mesh units, but really at some mesh vertexes the whole thing diverge..)
I wrote somewhere that giving a little gas flow is helpful?
Tips and tricks would be helpful
Lukas
Nagoya-u
sometimes I am lucky and I can run time step of 10E-7s, but most of the time it oscillates at 10E-8 or 10E-9, in bad cases it keeps on going at 10E-16s.. which of course never comes to an end.
I am happy when I can get up to 0,1 ms of plasma.. which is not always steady-state..
it is only Argon, and segregated (Field / Species) rather big 60000 mesh units..
but I still cannot find out why some models (parameters) run faster than the other..
to me it is very sensitive to the meshing.. but I do not know how I can evaluate this. ( not really dependant on the total amount of mesh units, but really at some mesh vertexes the whole thing diverge..)
I wrote somewhere that giving a little gas flow is helpful?
Tips and tricks would be helpful
Lukas
Nagoya-u
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
Hi Lukas,
Thanks for your kindly sharing of your own experience. The user must always be aware of the time step size, especially for capacitively coupled plasma reactors. Comsol makes use of the fully coupled numerical scheme which is numerically stable and a large time step can be used. However, a used large time step size does not mean that it is physically correct! Imagine that how can you resolve the plasma behavior with a sine wave (13.56 MHz) by a time step size like 5E-8 s since its period is around 7.3E-8 s? You may need at least a hundred or sometimes more than one thousand time step to get good result of the discharge since discharge is sensitive to the variation of applied voltage.
That is the point! If the simulation needs 100 time steps for one cycle with 1 time steps may take more than 10 s runtime and the whole simulation requires 10,000 cycles to reach quasi-steady state solution typically, you can count how much time the user may need to complete the simulation.
Hence, I tried all possibilities as I mentioned previously to reduce the runtime but unable to reach my goal. > <
Are those your tips and tricks available? Where I can find those tips and tricks? I am interested in any clue to reduce the runtime. Thanks in advance.
KM
Thanks for your kindly sharing of your own experience. The user must always be aware of the time step size, especially for capacitively coupled plasma reactors. Comsol makes use of the fully coupled numerical scheme which is numerically stable and a large time step can be used. However, a used large time step size does not mean that it is physically correct! Imagine that how can you resolve the plasma behavior with a sine wave (13.56 MHz) by a time step size like 5E-8 s since its period is around 7.3E-8 s? You may need at least a hundred or sometimes more than one thousand time step to get good result of the discharge since discharge is sensitive to the variation of applied voltage.
That is the point! If the simulation needs 100 time steps for one cycle with 1 time steps may take more than 10 s runtime and the whole simulation requires 10,000 cycles to reach quasi-steady state solution typically, you can count how much time the user may need to complete the simulation.
Hence, I tried all possibilities as I mentioned previously to reduce the runtime but unable to reach my goal. > <
Are those your tips and tricks available? Where I can find those tips and tricks? I am interested in any clue to reduce the runtime. Thanks in advance.
KM
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
Hello,
I made for now no progress.. I record each each time step as I set
10^{range(-8,6/1000,-2)}
into the times for the frequency-transient solving of an igniting DC plasma.
hoping to find out what is happening.
also I tried to set up a power supply with the circuit module, to follow also I/V behaviours (should follow Z too).
in the example for the microwave heated plasma they recommend a little convection of the gas..
tips and tricks ...
they are welcome
lukas
I made for now no progress.. I record each each time step as I set
10^{range(-8,6/1000,-2)}
into the times for the frequency-transient solving of an igniting DC plasma.
hoping to find out what is happening.
also I tried to set up a power supply with the circuit module, to follow also I/V behaviours (should follow Z too).
in the example for the microwave heated plasma they recommend a little convection of the gas..
tips and tricks ...
they are welcome
lukas