Add a short range repulsive potential as baseline potential

[PicoCentauri]

Problem

@HannaTuerk and @bananenpampe recently saw that for some of their systems atoms either cluster or their energy is smaller than one would expect if they are tiny. These are typical indications for holes in the potential. To get rid of them one can add more training data. But this can be very hard because sampling these small distances is very hard.

Possible solution

Another approach is adding a short-range repulsive potential as baseline potential and subtracting the energy and force before the actual training is started. This is very similar to the composition model.

The canonical candiate would be the Ziegler-Biersack-Littmark (ZBL) potential which is also implemented in MACE and GPUMD. @bananenpampe found also a Python implementation.

Implementation

The easiest approach is that we provide a function that computes the energy and forces based on the ZBL potential and each architecture can use this function if they like

Another idea I had, is that we provide these operations in a base class and architecture developers don't have to do it by themselves. Still, we already provide them the energies "corrected" by the composition energy and the short-range repulsion. Basically, they have to call super().forward() in their forward function. I don't know if this is a clever idea though...

[Luthaf]

Another idea I had, is that we provide these operations in a base class [...] Basically, they have to call super().forward() in their forward function.

This will not work, TorchScript does not support calling super().forward(). In general, having a ShortRangeZBL() module — like we have a LinearComposition module for isolated atoms energies — that contributors are using explicitly feels closer to the usual Torch way of doing things.

[PicoCentauri]

Fine for me as well.

Maybe I will on top create an architecture ShortRange that does the LinearComposition() plus ShortRangeZBL()

[HannaTuerk]

I implemented a ShortRangeZBL model which we can integrate. I am not sure where best to put it in the code, so it could discuss this. Otherwise I also talked to Sergey to put it directly in PET, which is also a possibility.
@Luthaf I did all the changes we talked about yesterday except for the requested_neighbor_lists, I not sure how to exactly implement that