Bayesian Optimization of Active Materials for Lithium-Ion Batteries

Abstract

The design of better active materials for lithium-ion batteries (LIBs) is crucial to satisfy the increasing demand of high performance batteries for portable electronics and electric vehicles. Currently, the development of new active materials is driven by physical experimentation and the designer’s intuition and expertise. During the development process, the designer interprets the experimental data to decide the next composition of the active material to be tested. After several trial-and-error iterations of data analysis and testing, promising active materials are discovered but after long development times (months or even years) and the evaluation of a large number of experiments. Bayesian global optimization (BGO) is an appealing alternative for the design of active materials for LIBs. BGO is a gradient-free optimization methodology to solve design problems that involve expensive black-box functions. An example of a black-box function is the prediction of the cycle life of LIBs. The cycle life cannot be predicted using a simple closed-form expression but only through the cycling performance test or a numerical simulation. BGO has two main components: a surrogate probabilistic model of the black-box function and an acquisition function that guides the optimization. This research employs BGO in the design of cathode active materials for LIB cells. The training data corresponds to the initial capacity and cycle life of five coin cells with different compositions of LiNixMn2 − xO4 in their cathode, where x is the content of Ni. BGO utilizes the experimental data to identify five new compositions that can produce cells with high initial capacity and\or large cycle life. The surrogate models of the initial capacity and cycle life are Gaussian Processes. The acquisition function is the constrained multi-objective expected improvement. The results show that BGO can identify high-performance active materials for LIBs. Designers can use the data generated during the optimization to decide the composition of the next batch of active materials to be tested, i.e., guide the physical experimentation.