Cost-aware sequential Bayesian optimal experimental design for lithium-ion battery life testing

Lithium-ion battery life testing commonly requires several rounds of expensive experiments to understand how design parameters like materials or operating procedures affect cell lifetime. To mitigate the cost of battery-life testing campaigns, we introduce a cost-aware Bayesian optimal experimental design (BOED) approach that sequentially samples the design space to gain as much information as possible about the mapping between design parameters and cell lifetime. In addition, testing until failure is only done when needed; otherwise, early-life failure prediction models are used. Two methods for cost-based decisions are proposed and compared: the first selects per-cell cycling time at the start of each experiment, and the second makes continual decisions on a cycle-by-cycle basis. Both strategies are demonstrated on five datasets with 1–6 design variables and compared to traditional experimental design strategies that represent performance baselines. Our proposed strategies are shown to outperform all baseline strategies considered, consistently reducing the total time required to reach equal confidence by at least ≈ 50% compared to the best baseline strategy. Practical applications of the framework are demonstrated using fast-charge optimization and formation protocol exploration problems, resulting in 3–7x reduction in experimental cost while discovering the same design principles.