My research group is working on some multiphase CFD simulations (non-reacting 2-phase flow).
We currently have a multiphase model set up in ANSYS with 1M node mesh and 0.02 time-step-size and it is taking about 1 minute per time step to converge. We are using a computer with 20 cores. Eulerian multiphase model. We estimate a 1-hour simulation time to take 125 days! That is a long time.
Would very much appreciate your advice on how to cut that time down / what you do for computing this kind of stuff! Thank you in advance for your input!
Regards,
Bryn
In my understanding ,20 cores and a small step size should be enough for fast computation even in case of Eulerian multiphase model. Are you using "parallel" as a processing option?
Unfortunately these things take time; and FLUENT is already one of the faster solvers out there. What you could try is to use NITA - non-iterative time advancement (see: http://www.cfd.com.au/cfd_conf15/PDFs/102FLE.pdf). It can significantly speed up your simulations, but in my experience, can also turn them unstable unfortunately. But it's worth a try considering the gain.
I agree with Cees.
I recommend you to use the Non-Iterative Time-Advancement (NITA) scheme. It helps to preserve overall time accuracy and gain a considerable time.
We gained a lot of time with validated simulations of dense industrial 3 phase Gas-Liquid-Solid flows. It was 3 to 4 times faster.
Best regards.
I think it takes time to run the multi-phase simulations. I am also having the same problem running the multi-phase simulations. I usually try to use "parallel" as a processing option, so that the simulation is a bit faster, but still it takes time to run such heavy simulations.
(May be I am misinterpreting your question. Do let me know if I get it wrong.)
If your mesh has 1 million nodes then it will definitely take a long time. According to me increase the mesh size to reduce the calculation time.
Bryn, one other question - what mesh type are you using? If you are using a tetrahedral mesh, you may gain by converting it to polyhedral. You can typically get good results at significantly lower mesh densities; although polymeshes need more operations per cell this is offset easily by the lower overall cell count.
Hey Bryn,
All suggestions above are great/legit. Which 2 phases are you considering? Which one is the dominant one?
We are considering a 2 phase model (non-reacting flows) with water at body temperature in both phases, but with slightly different density (e.g. it is slightly heavy water in the one phase).
Generally speaking i agree with Cees, both with regard to the fact that CMFD takes time and pros of using polyhedral mesh. Perhaps there is also some room to decrease simulation time by using variable time stepping.
Which two phase model do you use? Mixture, Eulearian, VOF?
You can also use the Variable Time Stepping method in Ansys Fluent. It changes the time step based on courant number. It may help a lot if your simulation is conversing in every time step.
Have you thought about the brute force approach: running on a supercomputer? I know with Star-ccm+, on an aerodynamics model, you can get 80% parallel efficiency down to 5,000 cells per core (superfast interconnect or 10,000 cells per core (standard supercomputer interconnect) . Hopefully you can get close to that with Fluent multiphase.
And why do you need to model 1h of physical time in the first place???What physics are you trying to capture? osmosis?
We are looking at a simulation time of 10 hours. It is for spread of a drug within the fluid around the brain (cerebrospinal fluid). This can take on the order of hours.
Why do you have to use Eulerian multiphase? Could you freeze the flow on an average flow solution and use a passive scalar approach for the drug dispersion?
- Badges
- Science topic