Structural descriptors evaluation for MoTa mechanical properties prediction with machine learning

Abstract

Abstract Considering all possible crystal structures is essential in computer simulations of alloy properties, but using density functional theory (DFT) is computationally impractical. To address this, four structural descriptors were evaluated using machine learning (ML) models to predict formation energy, elasticity and hardness of MoTa alloys. A total of 612 configurations were generated by the Clusters Approach to Statistical Mechanics software and their corresponding material properties were calculated by DFT. As input features of ML models, the CORR and SOAP performed best (R 2 > 0.90, some up to 0.99), followed by Atomic-centred Symmetry Functions, while Coulomb matrix performed worst. Furthermore, SOAP shows excellent performance in extrapolation for larger supercell structures of the MoTa alloy system and transfer learning for the MoNb alloy system.