How to Accelerate Running New Geometries for a Previously Trained Model

Hello!

I recently discovered the NVIDIA SimNet framework and am very excited to start testing it! I’m especially interested in cases like the aneurysm example. In particular, I would like to be able to train a model to predict fluid flow interacting with many different STL geometries. I read in the User Guide for the aneurysm example that Transfer Learning can be used to expedite the model training for new geometries but I have a a couple questions regarding this:

  1. Will I need new validation data for every additional STL geometry in order to perform Transfer Learning? If so, what benefit does SimNet offer compared to just running a classic CFD simulation for each new geometry?
  2. Is it even necessary to re-train the model for each new STL geometry? Can I instead train the model for one STL geometry, then use this model without re-training for several other STL geometries?

In general, what is the best approach to apply a previously trained model to a new version of the same problem (different boundary or initial conditions, same equations)? I’m assuming this is possible, as I think this capability is a primary use case for any simulation tool, but please let me know if this isn’t possible and I’m misunderstanding the capabilities of SimNet.

Thank you in advance for your time and help!

1 Like