Large-Scale Calculations by Integrating the Fragmentation Approach With Neural Network Potentials

Abstract

A fragmentation method is introduced to enable large-scale molecular simulations using neural network potentials (NNPs). The method partitions a system into cube-shaped fragments and reconstructs the total energy using a many-body expansion formalism with a distance-based cut-off approximation. Validation with Au, NaCl, diamond, H₂O, and graphite crystals demonstrated that including three-body interactions with 26 neighboring fragments reduces per-atom energy error to within 0.04 eV. This approach enables simulations of systems with up to 1 million atoms, surpassing conventional NNP limits. The scaling exponent for three-body calculations remains below 1.64, suggesting feasibility for even larger-scale applications.