About Modulus multi-physics coupling

Hello, I would like to ask about how Modulus does the multi-physics simulation. Imagine we have temperature and flow fields, which interact strongly. Is there one single NN trained for the two fields together or one NN for each field? If the latter case, how are the two NN coupled with each other? Thanks.

Hi @hitustckit

For a purely physics-driven training we have some examples that solve one way coupled heat sink problems (so the flow impacts the thermal field). I would say the most basic example that demonstrates this is the Scalar Transport example which has two models training together for flow and thermal fields.

Typically for more complex systems the flow field model is trained in isolation, fixed and then used to help train the model that solves the thermal field. An example of this is the FPGA example where we have both a flow and heat training script. Regardless we find having multiple models in mulit-physics systems is important for convergence since different physical quantities can have different structure and characteristics.

Hi. Thanks a lot for the quick reply. I read the documents of some related cases. Still some confusion:

  1. in traditional numerical solvers e.g. OpenFOAM, flow, and thermal fields are governed by continuity, momentum, and energy conservation equations. Those equations are solved together. If we use one-way coupling where the second does not affect the first, in some cases e.g. the natural flow where T strongly impacts D and V, does this algorithm still make sense?
  2. The coupled training of the “Scalar Transport example” does 2D geometry. Is that because this kind of “coupling” training requires a lot of computational resources so 3D is not practical?
  1. here is some condition of my case. I have two traditional solvers or codes, A is a thermal-hydraulic code solving water Density, Temperature, Velocity, and some dissolvable substances. B is a code that solves the thermal Power field. The D, T in A, and P in B are coupled tightly. I would like to implement a multi-physics NN which can predict coupled D, T, V, P using the data produced by A and B and also constrained by the Physical laws. Do you have some idea how I can implement it? Thanks a lot.
  1. Solving entirely for flow only first with one network followed by the thermal with another network is only possible if the flow variables are not temp. dependent (which is the case for heat sinks)
  2. When coupling becomes stronger, a nested solution approach is needed in the same network where you’ll do solve with one physics followed by another physics. This is done each iteration of the training so it’s coupled but it’ll lag by an iteration which in most cases will be fine
  1. If D & T in A and P in B are tightly coupled then you cannot use the data separately produced by each of A & B.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.