Inferring Interatomic Forcefields from Trajectories Using Graph Neural Network and Symbolic Regression

Classical molecular dynamics (MD)—the workhorse of computational material science—relies the accurate knowledge of the interatomic forcefields. In that regard, analytical forcefields are fast, physics-informed, and interpretable, but are hard to parameterize and often biased due to the choice of the underlying analytical form. On the other hand, machine-learned forcefields are non-biased, but are more computationally expensive and tend to offer poor generalizability to new phases or conditions. Here, we present a new approach that is able to infer analytical interatomic forcefields from trajectories produced by Density Functional Theory (DFT). Our approach is based on fitting DFT results by a message-passing graph neural network (GNN), which is then simplified into an analytical forcefield using symbolic regression—without any assumption regarding the nature of the analytical form. We show that the forcefields that are discovered offer an accuracy that matches that of machine-learned forcefields, at a fraction of the computing cost, and feature improved generalizability.