Gradient of forces with respect to positions

[sigbjobo]

Hi, I am working on a project where I need to compute the gradient forces with respect to particle coordinates. Right now, I am thinking to do this numerically, but it is less elegant than doing it through directly.

Is there a simple way to achieve this with the current version of nequip/allegro? And if not, is there a simple way to implement this?

[Linux-cpp-lisp]

Hi @sigbjobo ,

I think this should be addressed by #292 (comment). Please note that since that post the functorch / torch.autograd.functional APIs have evolved and matured somewhat, and there may be better / faster ways to do it, such as trying with vectorize=True.

I'd still be happy to accept a PR adding this and don't think it should be too much work. (Fair warning, 2nd derivatives in PyTorch were, at least last time I checked, not particularly fast.)

[sigbjobo]

Thanks, this sounds promising! I will definitely submit a PR if we manage to do it.

[Linux-cpp-lisp]

Great! Yes, it should be fairly straightforward I hope.