Adapting NequIP and Allegro for Classification Tasks

[ankur56]

Hi everyone,

I'm working on adapting NequIP and Allegro for classification tasks. Given that these architectures are primarily designed for energy and force predictions in atomistic simulations, I have a few questions about modifying them for classification:

1. Would changing the loss function (e.g., to cross-entropy) and providing discrete target labels be sufficient for basic classification, or would it require more complex modifications to the architecture?
2. For the force-related components, could I simply set the coefficient of the force term to zero if it's not needed for classification, or would this approach cause issues?
3. Are there any specific considerations or challenges I should be aware of when attempting to use these models for classification instead of regression?

Any guidance or insights from those experienced with these models would be greatly appreciated!