Allowing the "features" output to have rotational information.

[pfebrer]

Until now, things have been designed for PET, whose features don't have a predefined rotational behavior. However, equivariant models are being added into metatrain (MACE and PhACE), whose features have a well defined rotational behavior.

When running an AtomisticModel with the consistency check turned on, the equivariant features cause an error because this possibility is not yet contemplated.

We should somehow make AtomisticModel aware that features can be equivariant. The difficulty in doing this (as I understand) is what to do with interfaces to other codes that for now also don't support these equivariant features.

[Luthaf]

So the "features" output was designed for chemiscope and PLUMED, with the idea that these can be used for visualization or as ML collective variables.

I'm not 100% sure what the use case is for equivariant features here, other than most models in metatrain have a "features" output so MACE should have one as well =). Do you know where such feature would be used?

[pfebrer]

Sure, if you want to add a head on top of the metatrain's implementation of any equivariant model, you need to know what are the "irreps" that these features are representing.

[Luthaf]

But this is an internal thing inside the model, not an output that needs to be exposed outside of the model, right?

[pfebrer]

At some point there was a tutorial in metatrain showing how to use the features of PET programmatically, I can't find it now