An accurate machine-learning calculator for optimization of Li-ion battery cathodes

Area of Science:

• Materials Science
• Computational Chemistry
• Electrochemistry

Background:

• Improving battery performance is crucial for electrifying transportation and aviation.
• LiNi_xMn_yCo_(1-x-y)O₂ cathode materials show promise, but computational expense hinders design.
• Predicting open circuit voltage requires evaluating stability across complex phase spaces.

Purpose of the Study:

• Develop accurate and efficient machine-learning potentials for battery cathode materials.
• Enable rapid design optimization of LiNi_xMn_yCo_(1-x-y)O₂ materials.
• Validate machine-learned potentials against experimental battery performance.

Main Methods:

• Utilized atom-centered symmetry functions and neural networks for machine-learning potentials.
• Trained models on density functional theory (DFT) calculations.
• Employed Bayesian optimization for hyperparameter tuning and grand canonical Monte Carlo simulations.

Main Results:

• Achieved high prediction accuracy for energy (3.7 meV/atom) and forces (0.13 eV/Å).
• Accurately predicted thermodynamic properties, including Gibbs free energy and entropy.
• Simulated Li-vacancy ordering and reproduced experimental voltage profiles.

Conclusions:

• Machine-learned DFT surrogates offer a computationally efficient approach for battery materials optimization.
• This method accelerates the discovery and design of advanced cathode materials.
• Demonstrated the potential for rapid design optimization in complex material systems.