Physics-Based Long-Range Equivariant Interatomic Potentials

Machine learning interatomic potentials offer the ability to perform large-scale atomistic simulations at a fraction of the cost of first-principles density functional theory simulations. However, the favorable linear scaling of machine learning interatomic potentials comes at the expense of reduced accuracy due to neglect of long-range interactions. These interactions can become significant or may be required to simulate materials and molecules with different charge states. In this work, we address this limitation by incorporating long-range interactions into a neural equivariant interatomic potential by determining the atomic charge distribution using a charge equilibration scheme. The scheme ensures that charges are distributed on atoms according to their local electronegativities, enabling accurate representation of charge transfer and polarization effects. Our model architecture consists of two key components: one to predict the charge distribution within the system and the other to compute the total energies and forces after incorporating these charges. We evaluate our model on a diverse set of benchmark datasets from the literature, including both periodic and non-periodic systems. The proposed long-range model outperforms local descriptor-based interatomic potentials in predicting energies and forces. Moreover, it demonstrates data efficiency, requiring fewer training samples to reach accuracy levels comparable to or better than existing potentials. This potential will enable more accurate and efficient molecular dynamics simulations of systems with long-range interactions, including those with different charge states or involving charge transfer.