Hello,
I would like to solve a high density fluid flow(e.g. water) that is 3D and unsteady.
I’m trying the taylor-green example with water property(nu=0.000001, rho=1000) but the reduction of train loss is much slower than one in the original setting(nu=0.002, rho=1.0) and the flow doesn’t evolve with time at all.
I used 22.07 for taylor-green and 22.03 for some other problem with high density flow.
All of their results are not well.
If you have idea to solve high density fluid flow, please let me know
Regards,
Hi @yokoi.toshiaki
Success in physics-informed learning greatly depends on proper scaling of the different quantities present in the physical system. For example the domain may be adjusted for a higher viscosity, etc. There’s on info on this here. Without a proper scaling of the physical system, the machine learning will be nearly impossible. This can include material properties, extremes rarely work without a lot of work.
I have not personally tried Taylor Green on the parameters, but the very small viscosity is effectively removing the influence of the diffusion term. This could be causing the model to fall into a local minima based on the initial condition that it cannot escape for learning any dynamics in time. This is just speculation of course.
Regardless, by reducing your dynamic viscosity and increasing density will make the Reynolds number very large entering turbulence, which will be extremely challenging (not possible presently) for PINNs with zero data. Hence the reason PINNs is typically used for lower Reynolds number flows.
Thanks for your detailed reply.
As you mention the fundamental problem depends on scaling and optimization I think too.
I’m going to investigate it more.
Thanks,
1 Like
Hello,
Is there any progress made since then?
I’m facing the exact same issue, but didn’t find a way to solve it yet, even with scaling and adjusting my network and weighting.
It would be really helpful to know if anyone has solved this problem and how.
Thanks a lot in advance.