Fine-tuned models with TorchSim

[dominicvarghese]

Hi everyone,

When loading different MLIPs via calculator(), is there any change required depending on whether the model is pre-trained or fine-tuned?

In particular, if a model was fine-tuned using multi-head training, how do we ensure that the correct head (e.g., the default or fine-tuned head) is selected during inference? Is there any additional argument that needs to be specified in the calculator() call?

Thanks
Dominic

[CompRhys]

Things like the selection of model heads are very model specific and the complexities of handling this is part of our push to an external model posture (see #120) such that individual model authors can expose whatever kwargs they like provided that they keep to the API contract.

Which model are you trying to use here as that informs the extent to which we might be able to offer any help.

[dominicvarghese]

I was interested in looking at the fine-tuned MACE and PET architectures.

So does that mean that the universal models available (in the Matbench discovery) can be accelerated with TorchSim?

[CompRhys]

A torch-sim ModelInterface can be made for any model, including those that have submitted results to MBD, whether an interface exists for you to use is a different question. Our hope is that as more people use MLIPs then model authors will begin to implement TS interfaces as well as ASE calculators when releasing models.

Both MACE and PET should have interfaces here via the mace and metatomic files but they might not expose the kwargs you need for multi-head models and head selection.