choderalab / modelforge

Infrastructure to implement and train NNPs
https://modelforge.readthedocs.io/en/latest/
MIT License
9 stars 4 forks source link

Molecular inference in training SpookyNet #164

Open ArnNag opened 2 weeks ago

ArnNag commented 2 weeks ago

SpookyNet requires information about which atoms belong to the same molecule to model distance-independent "quantum nonlocality and electron delocalization." In addition to tracking which record each atom was obtained from, which is done by enumerating pairs within records in PairList, we also have to take into account that SPICE records can have multiple molecules (for example, solvated amino acids and water clusters). This is somewhat analogous to the need to infer bonds to apply Monte Carlo methods in simulation (@chrisiacovella can explain this better).