Message Passing Network with Iterative Charge Equilibration (MPNICE)
While the literature has tended towards larger models with increasing numbers of parameters and equivariant featurization, we’ve found that a relatively simple invariant message passing architecture coupled with iterative charge equilibration and long-range coulomb interactions results in similar or greater accuracy while keeping inference costs an order of magnitude lower than comparable MLFFs. Additionally, by incorporating charge equilibration, MPNICE can simultaneously generalize to systems with distinct charge states, and even directly access higher order response properties to external electric fields.