Benchmarking equivariant machine learning interatomic potentials with long-range interactions

Incorporating long-range electrostatics into machine learning interatomic potentials (MLIPs) is essential for modeling systems with long-range charge transfer. However, there is a lack of benchmarking data for long-range MLIPs and guidance on choosing hyperparameters related to charge prediction. Here, we systematically benchmark a long-range MLIP, NequIP-LR [1], on diverse datasets and compare its performance using different hyperparameters. We further investigate whether it is possible to obtain accurate predictions without explicit charge equilibration within a long-range MLIP, which is computationally expensive. This MLIP, NequIP-noQeQ, predicts charges directly as atomic properties, in contrast to NequIP-LR. The straightforward way of learning charges from energies and forces feature enables NequIP-noQeQ to be more computationally efficient than NequIP-LR when datasets have no ground truth charges. Overall, we establish a rigorous framework for evaluating and improving long-range MLIPs for materials with significant electrostatic effects.

[1] M. U. Maruf et al., J. Phys. Chem. Lett. 2025, 16, 35, 9078–9087.