Machine learning interatomic potentials in engineering perspective for developing cathode materials

Abstract

Efforts to develop advanced battery materials have been ongoing for decades, yet progress in cathode materials used in commercial applications has stalled. With recent advancements in artificial intelligence, machine learning has been increasingly utilized in various fields, including material development. Attempts to replace ab initio calculations with machine learning interatomic potentials (MLIPs), which are much faster than density functional theory (DFT) calculations, are becoming increasingly sophisticated and prevalent. Currently, MLIPs focus on accurately predicting material properties; however, research comparing their practicality and applicability to observations and first-principles calculations from an engineering perspective are limited. Therefore, this study used two of the most recently developed MLIPs, namely Crystal Hamiltonian Graph neural Network (CHGNet) and MatErials 3-body Graph Network (M3GNet), to validate the engineering application of MLIPs from thermodynamic and structural perspectives. The assessment was conducted using the olivine structure of LiFePO₄ with Mn substitution and rock salt structures of Li-rich Li₂TiO₃ and Li₂TiS₃, and the results were compared with Perdew–Burke–Ernzerhof (PBE) pseudo-potentials. In the absence of vacancies, the MLIPs accurately replicated thermodynamic phase stability and ionic configuration, exhibiting values comparable to those observed with PBE. As the Li vacancy content increased during model charging, the discrepancy in the PBE results became more pronounced. A comparative analysis of the MLIPs and DFT demonstrated that MLIPs can be used to calculate the thermodynamics and lattice constants of non-defect crystals with a high degree of similarity to the results obtained from DFT. Owing to their rapid computational capabilities, MLIPs are expected to be beneficial for screening potential candidates for the development of novel materials.