SG-NNP: Species-separated Gaussian Neural Network Potential with Linear Elemental Scaling and Optimized Dimensions for Multi-component Materials

TL;DR

This paper introduces SG-NNP, a robust, scalable Gaussian neural network potential for multi-component materials that improves accuracy and efficiency in long-term, large-scale simulations by using optimized, species-separated descriptors.

Contribution

The paper develops a novel set of Gaussian descriptors that scale linearly with atoms and enhance robustness for multi-component environments, applied in SG-NNPs.

Findings

SG-NNP outperforms traditional potentials in energy and force predictions

The method scales linearly with the number of atoms

Improves descriptor performance for complex multi-element systems

Abstract

Accurate simulations of materials at long-time and large-length scales have increasingly been enabled by Machine-learned Interatomic Potentials (MLIPs). There have been increasing interest on improving the robustness of such models. To this end, we engineer a novel set of Gaussian-type descriptors that scale linearly with the number of atoms, reduce informational degeneracy for multi-component atomic environments and apply them in Species-separated Gaussian Neural Network Potentials (SG-NNPs). The robustness of our method was tested by analyzing the impact of various design choices and hyperparameters on Molybdenum (Mo) SG-NNP performance during training and inference/simulation. With less dimensions, SG-NNPs are shown to have superior atomic forces and total energy predictions than other traditional and ML descriptor-based interatomic potentials on diverse set of materials - Ni, Cu, Li, Mo, Si, Ge, NiMo, Li3N and NbMoTaW. From the obtained results we can observe that the proposed method improves the performance of atomic descriptors of complex environments with multiple species.