Foundational Machine Learning Interatomic Potential to Study Li-Ion Battery Cathode Phase Transformation with Charge Transfer

Large-scale simulations with complex electron interactions remain one of the greatest challenges for atomistic modeling of electrochemical systems. Our recent work, Crystal Hamiltonian Graph Neural Network (CHGNet), presents foundational graph-neural-network-based machine-learning interatomic potential (MLIP). The inclusion of charge information by magnetic moment prediction enables CHGNet to better describe both atomic and electronic degrees of freedom, which opens the possibility for a better understanding of electron-redox coupled behaviors that are essential for applications such as battery cathode materials for energy storage. We demonstrate the use of CHGNet MLIP in electrochemistry by modeling in detail the process by which orthorhombic LixMnO2 transforms to a spinel-like structure. This transformation is facilitated by the valence-dependent Mn migration and charge disproportionation which can only be modeled accurately with an MLIP that captures the very different chemical behavior of transition metals’ different valence states. Moreover, we show the Mn-rich disordered rocksalts (DRX) exhibit similar phase transformation to a spinel-like phase, leading to improved capacity and Li-ion transport kinetics. Our study demonstrates critical atomic-level insights into Li-ion cathode electrochemistry, which became possible with charge-informed long-time simulations at scale enabled by modern AI methods.