Modeling Charged Atoms for Atomistic Dynamics, using Equivariant GNNs

Graph Neural Networks have been quiet successful in modeling inter-particle interactions at quantum level for datasets carrying force-fields and atomic charge, detecting efficacy of a drug compounds, preventing further harmful chemical reactions such as drugs targeting cancerous cells, preventing tumor growth (protein ligand interactions). There are various other applications of this approach where one can model inter graph interactions, among nodes belonging to different graphs, using cutoff-radius based sphere as receptive field such as recommendation system taking into account of user’s interactions with influencers on other social networks, interactions among various sets of swarm drones, spread of viruses among multiple species of birds, based on their interactions across flight paths etc... Dataset such as SPICE [3], have information about cutoff-radius embed in them. Selecting right cutoff-radius is key parameter for getting reasonable outcome. Next is AGGR function as authors have notices that charge should not be approximated using SUM but MAX to get current bond-length measure at atomic level. For other domains, prior domain knowledge should guide these decisions. We present a e3nn, nequip, Allegro based GNN with slight modification for SPICE dataset carrying charge, to detect right bond length even with short training of 100 epochs. Ideas presented here can be applied to other node interactions belonging to various graphs. Here we show how we detected ammonium ion N-H bond length, where one N-H bond out of four carried charge. We have a Jupyter notebook for readers to try our approach in less than ten minutes.