How to use the moving time window method when solving a convective heat transfer problem


I need to solve a transient heat transfer problem in which the flow field and temperature field are coupled. In the Taylor-Green vortex decay example we have just Navier-Stokes eq. and a flow_net. How can I extend the moving time window method for problems in which there is a heat_net as well as a flow_net?

Hi @z.hashemi986

Theoretically this is possible, but even taylor-green with this approach is not an easy problem. A transient coupled heat transfer problem is something that is even more complex than we have done, so unless you are very familiar with PINNs, I would recommend starting simple and building up. Its unknown how well the moving-time window approach generalizes to other problems / physics.

A good research problem. Best of luck.

Thanks for your reply.
By the way, I’m using GPU TRX4090 Ti, but I cannot run the “industrial heat sink” example and I get an “out of memory” runtime error.
Except for A100, which GPU cards do you recommend for PINN using Modulus (for a transient coupled heat and fluid flow may involve phase change)?
Do you recommend any item from these models: A6000 or A40 or A16?

Hi @z.hashemi986

For some of these more complex problems, GPUs with more VRAM are recommended (such as A100 or V100s). Generally with deep learning more VRAM you have the better. But for many PINNs problems , other GPUs will work just fine. A 4090 has quite a bit of VRAM so maybe adjusting some hyper-parameters may fix it:

Thank you so much for the help.