Electrolyte recommender system for batteries using ensemble Bayesian optimization

Abstract

Bayesian optimization (BO) is a useful technique for optimizing unknown target functions that require computationally demanding evaluations and have noisy outputs. In addition, BO can be easily extended to parallel experiments. Ensemble BO, involving multiple BO algorithms, is suitable for such settings and resolves some of the issues associated with classical BO such as mismatches between the unknown target and surrogate functions. In ensemble BO, low-performance results are actively gathered for the purpose of training BO algorithms and for use in analyzing target phenomena of interest. In this study, we apply ensemble BO to a real experimental system: a system to recommend the optimal electrolytes for Li-ion batteries. The purpose of this recommender system is to determine the best solution (solvents and salt) ratios for the electrolyte to maximize ionic conductivity. The system has a large search space in terms of the possible combinations of solution ratios, and must deal with noisy observations. The proposed electrolyte recommender system, based on ensemble BO, was experimentally demonstrated to be effective for a battery electrolyte with four solutions. Furthermore, based on the solution ratios recommended by the system, a useful solution-mixing rule for maximizing ionic conductivity was obtained.