Graph Neural Network-Based Interatomic Potential Calculations Combined With Grand Canonical Monte Carlo Simulation to Predict the Electrochemical Potential Profile: A Model Study Using Spinel-Type Titanates

Density functional theory (DFT) is an effective approach for the in silico design of battery materials; however, DFT calculations are usually performed at 0 K, and the resulting electrochemical potential profiles are stepwise and do not resemble the continuous profiles obtained experimentally. Therefore, the estimation of temperature effects on electrochemical potential profiles remains challenging. Recently, theoretical calculations based on graph-neural-network-based interatomic potentials (GNNPs) have garnered attention because of their computational speed, enabling energy estimation several orders of magnitude faster than DFT. However, because GNNPs are trained on DFT data, simulating and studying continuous energy profiles using only GNNPs has remained impossible. Herein, GNNP calculations were performed to predict electrochemical potential profiles by combining grand canonical Monte Carlo simulations, using a spinel-type titanate as the model compound. GNNPs trained solely on Materials Project data without dispersion force correction did not reproduce experimental results even qualitatively, whereas GNNP with the correction reproduced experimental results. Furthermore, a GNNP trained on both stable and unstable structures could be utilised for electrochemical potential profile prediction. These results indicate that handling unstable structures is important for capturing thermodynamic properties, and that inadequate sampling of unstable structures can be mitigated by dispersion force corrections.