Autonomous Discovery of Polymer Electrolyte Formulations with Warm-Start Batch Bayesian Optimization

Abstract

Solid polymer electrolytes are a promising class of materials to enable next-generation Li-based batteries. They offer highly tunable properties, scalable processing conditions, and increased safety. However, current solid polymer electrolytes do not have sufficient ionic conductivity for room-temperature battery applications. The discovery of novel polymers and the optimization of polymer-salt formulations with high ionic conductivity are critical bottlenecks in developing new polymer-based batteries. Programmable laboratories driven by machine learning algorithms have been proposed to power accelerated discovery cycles. Here we demonstrate a closed-loop, machine-learning driven Bayesian optimization pipeline for optimizing a dry polymer electrolyte composed of poly(ϵ-caprolactone) (PCL) electrolyte with one of 18 lithium salts. We use previously collected literature data to warm-start our optimization and achieve high performance while following through with a novel high-exploration batch-based sampling method. Formulations chosen by the sampling method were mixed, cast, dried, and characterized on an autonomous high-throughput polymer electrolyte platform. After five batches of optimization conducted in just over a month, we discovered formulations with ionic conductivity that were on par with top-performing poly(ethylene oxide) electrolytes, the standard of the field.