Triplet Pruning and Spatial Decomposition in M3GNet: Accuracy and Memory Trade-offs in Scaling Graph Neural Network Potentials

Abstract

Machine-learning interatomic potentials of several functional forms now support large-scale atomistic simulation.Within this landscape, graph neural network (GNN) potentials such as M3GNet offer transferable message-passing models, but their application to very large systems remains limited by graph construction, message passing, and differentiable force-evaluation costs.In M3GNet, the explicit angular pathway is represented through line-graph triplets, whose construction can dominate inference cost.Here we examine the scalability of M3GNet molecular dynamics via inference-time triplet pruning.We first show that an apparent insensitivity to triplet removal in inspected MatGL implementations originates from an incorrect parent-bond index mapping; after correction, removing triplets substantially increases static energy, force, stress, and property-prediction errors.In short molecular-dynamics trajectories, however, the structural response is strongly system dependent, with some systems showing modest radial-distribution-function changes despite a threefold reduction in inference cost.We then implement and evaluate an owned-core/halo spatial chunking scheme that bounds activation memory during differentiable force evaluation and enables M3GNet simulations of million-atom alloy cells on GPU hardware.These results show that triplet pruning can provide useful acceleration when validated for the target observable, while spatial decomposition offers a more general route to extending GNN potentials to large atomistic systems.