My research is looking into a generalized approach for symmetry extraction from ML models. The idea is to find generators of the symmetry groups and use them to generate new data while satisfying multiple desired equivariance condition liked to the properties defining the symmetry groups. In the first place, I also took a look into graph decomposition using neural networks to identify individual molecules in atomistic simulations. This approach was furthermore used to make predictions on internal bonding properties and has the possibility to be extended to external interactions with other molecules and atoms including VdW-forces / hydrogen bonding and numerous other interaction.
However, we are still struggling to find big datasets to train in these regimes and need to fine tune the model further.
If you have any suggestions or questions I would appreciate if you come by my office or let me know by mail.