Developing Machine Learning Force Fields of High Entropy Alloys with Deep Neural Networks*

Computational simulation has become an invaluable tool in modern materials physics. Broadly, materials modeling falls into two categories: highly accurate but computationally intensive ab initio methods like density functional theory (DFT), and time-efficient but contextually limited methods such as classical molecular dynamics (MD). The need for a parameterized potential has long constrained the accuracy of classical MD. However, the recent rise of deep learning has opened a new path to combine quantum ab initio accuracy with classical MD efficiency. In this work, we investigate the ability to develop a DeePMD potential using DFT data to accurately model a wide range of metallic compounds—from binary alloys to emerging high-entropy alloys—under a single unified potential. The results offer insights and suggested practices for developing deep learning potentials in future studies of compositionally complex, multi-component materials.