Multi-Objective Bayesian Optimization for Materials Discovery with Neural Network Potential—An Application to Li-Ion Battery Cathode Material

Cathode materials are the key factors of Li-ion batteries and many properties must be considered. Traditionally, it was necessary to narrow down material candidates in advance using domain-specific knowledge as represented by the filtering approach due to the prohibitively expensive density functional theory (DFT) calculations. On the other hand, a fast approximation for DFT calculations such as a neural network potential (NNP) has enabled a large-scale joint optimization of multiple properties for materials discovery rather than the filtering approach.<br/>Our primary aim is to demonstrate that it is possible to successfully draw material candidates with better trade-offs over existing materials in terms of a user-provided set of properties to optimize when we combine multi-objective Bayesian optimization (BO) with an NNP. Furthermore, we investigate search space design that enables more efficient materials discovery. To illustrate an example, some properties of Li-ion battery cathode materials will be optimized.<br/>Our materials search was conducted using our developed universal NNP called PFP, and Optuna, a flexible BO library mainly developed by us that easily realizes multi-objective optimization equipped with rich features such as a web dashboard and a large-scale distributed optimization. At each iteration, Optuna wisely learns to replace the transition metal of existing cathode material, e.g., LiFePO4 and LiCoO2, with other transition metals such as Mn and Ni based on obtained properties, i.e., voltage, capacity per mass, and cost, during an optimization. Additionally, we tried various different decision (or search) spaces for the metal substitution to study the search efficiency depending on search space designs.<br/>The numerical experiments using PFP and Optuna with 50 GPUs allowed us to find 200+ Pareto solutions out of 10,000 structure evaluations in less than 10 hours. Furthermore, we confirmed that the search space design is a key factor in determining the efficiency of the search. Although the set of properties used in our work is very limited, our results exhibited a possibility that a well-crafted property design enables us to find better material candidates in combination with multi-objective optimization and our search space design.