Autonomous Electrolyte Discovery for Batteries with Experimentally Informed Bayesian Optimization

Abstract

An autonomous battery electrolyte experimental platform capable of mixing multi-component electrolyte systems and characterizing the transport and electrochemical properties in a high-throughput manner is disclosed. A Bayesian optimization software package found novel electrolyte compositions through optimization of the high-dimensional electrolyte design space over key design objectives like electrochemical stability and conductivity.Electrolyte optimization is difficult because 1) electrolyte evaluation is expensive and takes time, and 2) the space of possible electrolytes is expansive, formed by many possible choices for solvents (often in ternary blends), salts (often in binary mixtures), and trace additives. Bayesian optimization methods are well suited for the optimization of high-dimensional functions with costly evaluations, often producing an efficient design-of-experiments to converge on multi-objective optimal formulations in few experiments. To expedite the optimization over the expansive design space, theoretical predictions of electrolyte properties via the Advanced Electrolyte Model were utilized as “priors” in the statistical model.Implementation of the novel experimental platform was carried out by two novel test stands developed to automate the mixing and characterization of electrolytes: Otto (for aqueous systems) and Clio (for aprotic/organic systems). The test stands characterized the ionic conductivity and electrochemical stability of electrolyte systems, featuring a four-electrode conductivity probe, pH meter, and a flow-through three-electrode cell and potentiostat. Clio also integrated electrochemically active electrodes for optimization of electrolyte/electrode systems. The active electrode systems used were common functional ceramic metal oxides.A software orchestration and data layer linked the test-stands to human experimenters and machine-learning packages through a web-services architecture; all experiment data and meta-data is saved in a database. Additional out-of-the-loop characterization was conducted on cathodic systems to validate composition, structure, and oxidation state.The aqueous design space consisted of aqueous blends of lithium and sodium salts, including nitrates, sulfates, and other commonly-used battery salts. High-concentration aqueous electrolyte candidates were discovered by optimizing of electrochemical voltage stability and conductivity, including low-cost, high-performing alternatives to known but costly aqueous electrolytes (e.g. LiTFSI). Clio’s design space includes blends of both aprotic organic solvents and solutes in additional to various compositions of electrochemically active electrodes. The test-stands are demonstrated to be significantly faster than common human experimentation techniques, converging on novel, optimized electrolyte mixtures in mere hours or days of experimentation.