Linear Scaling Calculation of Atomic Forces and Energies with Machine Learning Local Density Matrix

TL;DR

This paper introduces HamGNN-DM, a machine learning model that predicts atomic forces and energies with linear scaling, enabling efficient and accurate large-scale molecular dynamics simulations at DFT-level precision.

Contribution

The paper presents a novel machine learning approach using local density matrices for linear-scaling atomic force and energy predictions in molecular dynamics.

Findings

Achieves DFT-level accuracy in force predictions across various systems.

Operates with O(n) time complexity suitable for large systems.

Provides electronic structure information during simulations.

Abstract

Accurately calculating energies and atomic forces with linear-scaling methods is a crucial approach to accelerating and improving molecular dynamics simulations. In this paper, we introduce HamGNN-DM, a machine learning model designed to predict atomic forces and energies using local density matrices in molecular dynamics simulations. This approach achieves efficient predictions with a time complexity of O(n), making it highly suitable for large-scale systems. Experiments in different systems demonstrate that HamGNN-DM achieves DFT-level precision in predicting the atomic forces in different system sizes, which is vital for the molecular dynamics. Furthermore, this method provides valuable electronic structure information throughout the dynamics and exhibits robust performance.