Including periodic translations invariance in the SchNet implementation.

[fedeotto]

I'm wondering if it is straightforward to modify the SchNet torch geometric implementation by including periodic conditions, as described in the paper? I guess the actual implementation relies on an earlier version of the paper..?

[rusty1s]

I haven't looked into that for a while, but I don't think it should be complicated to add. The current implementation replicates the trained benchmark-models of schnetpack from here.

[fedeotto]

Hi Matthias and thanks for the kind answer.

Looking at the implementation of this paper, I've found that these people have obtained that by modifying radius_graph function in the case of DimeNet++. I still have to figure out exactly what they do though.