Machine Learning, Molecular Dynamics and Excitons II

Machine learning potentials (MLPs) hold great promise for large-scale atomistic simulations, which are essential for addressing a wide range of significant problems in materials science, chemistry, and molecular biology. Despite the rapid advancement of MLPs, rigorous evaluations of their performance in actual simulations remain limited. In this work, we showcase the application of MLPs, specifically those based on graph neural networks (GNNs), to accelerate large-scale Monte Carlo simulations. We provide a comprehensive assessment of the accuracy, generalizability, and efficiency of these machine learning models, comparing them to traditional methods such as density functional theory (DFT) and empirical potentials. Our results show that GNN models achieve speeds approximately $10^6$ times faster than DFT, handling systems over 1000 times larger while maintaining chemical accuracy. Additionally, we applied this approach to study the order-disorder transition in high-entropy alloys, shedding light on the origins of their exceptional strength and ductility.