PDE Based Training of New Models in Modulus Core


I’m interested in trying out the graph and recurrent network models, introduced in Modulus Core, with a PDE based loss function. Is there currently a way to do this? Modulus Sym currently doesn’t have those models while Modulus Core doesn’t appear to be set up for custom PDE-based loss functions.


Hi @rohan.patel

That is correct. Modulus (Core) is a more PyTorch developer orientated tool kit while Modulus-Sym provides an abstract framework with the symbolic loss representation. Presently PDE based loss calculations (with automatic gradient calculations) is something limited to Modulus-Sym, which does not support RNNs or graph NNs.

Models in Modulus Core presently require you the manually write PDE loss functions in standard PyTorch methods.

The RNNs and GNNs in Modulus core were developed with data-driven applications in mind, thus not fully compatible with Modulus Symbolic at the moment.