Construction and Application of Machine Learning Force Fields in 9-element High Entropy Alloys Made of Non-Noble Metals.

High entropy alloys (HEAs), composed of 5 or more elements in equal proportions, display unique properties due to the "cocktail effect." Atomistic simulations using machine learning force fields (MLFFs) are gaining attention because the vast combinations of elements make experimental research challenging; However, constructing MLFF with 6 or more elements is difficult due to many interaction patterns. In this study, we constructed and evaluated various MLFF models for a TiZrNbMoCrCoNiMnFe 9-element HEA [1] from density functional theory (DFT)-based molecular dynamics (MD) simulations. GAP-SOAP achieved the lowest force root mean squared error (RMSE) (0.359 eV/Å) compared with sGDML (0.895 eV/Å), n2p2 (0.437 eV/Å), and SchNet (0.375 eV/Å). The GAP-SOAP’s performance stems from its use of spherical harmonics-based descriptor and Gaussian kernel, resulting in fewer functions compared to other MLFF models, particularly those based on neural networks. Radial distribution functions (RDFs) calculated from MD simulation at 300 K and 3000 K using GAP-SOAP closely matched those from DFT-MD, validating its reliability for structural analysis. In addition, the RDFs showed no significant change when the number of atoms in the unit cell increased from 216 to 1728 (8 times). This consistency suggests that GAP-SOAP remained accurate in larger systems.



[1] A. A. H. Tajuddin et al., Adv. Mater, 35, 2207466 (2023).