Development of machine-learned force fields for lithium-manganese-oxide (Li-Mn-O) spinel. Density functional theory accuracy at a lower computational cost

The on-the-fly machine learning implemented in the VASP code was utilized to generate interatomic potentials for LiMn2O4 spinel. The potentials will be further expanded to incorporate the Li2MnO3 layered character to enable accurate large-scale atomistic simulations of layered-spinel composite structures at practical computational cost. Li2MnO3 is incorporated in LiMn2O4 structures, which are plagued by capacity fading due to the loss of manganese caused by irreversible structural changes, to improve the electrochemical performance of this material. As such, the generated MLFFs in the current study are a significant step in that direction. The MLFFs were able to reproduce the DFT data with a percentage difference of less than 2%, showing impressive accuracy. The generated MLFFs will be further expanded and improved in a future study.