Architectures on nvidia modulus for data-based model and evaluation of the trained neural network?

Physics Simulation Modulus Physics-ML Model Framework Technical Support (Modulus Only)
1

I want to train a neural network with only data, the data I want to use for training correspond to solutions obtained from CFD simulations in ansys fluent in a 3d domain. What model should I use for this purpose? Fourier neural operator is the indicated option? Are there other architectures in nvidia modulus that can be useful for training only based on data? And finally, after training, using the modulus extension for Omniverse, is it possible to evaluate this already trained neural network with new input parameters?

2

@matiyanez

Thanks for your interest in Modulus, some responses to your questions are below:

What model should I use for this purpose? Fourier neural operator is the indicated option?

Many models can be used. If you want a continuous function then a fully connected, Fourier neural network would work. If your data is on a structured grid then FNO, AFNO or convolutional (pix2pix, super-res) models can work.

Are there other architectures in nvidia modulus that can be useful for training only based on data?

FNO / pix2pix / super-res can all work well for structured data. The fully-connected models can work well too if your data is unstructured. Its very problem dependent (based on data, system complexity, etc.).

And finally, after training, using the modulus extension for Omniverse, is it possible to evaluate this already trained neural network with new input parameters?

Yes but the OV extension is in a beta state with limited support and documentation. This code can be quite challenging to develop inside omniverse. However, there are several examples that you can build off of if desired.

1 Like