(435e) Data-Driven Electrolyte Design for Lithium Metal Batteries

Enhancing Coulombic efficiency (CE) plays a pivotal role in facilitating the adoption of high-energy-density lithium metal batteries. While liquid electrolyte engineering has emerged as a promising strategy for improving CE, its inherent complexity makes performance prediction and electrolyte design challenging. In this two-part presentation, we introduce machine learning methods and workflows that enable the prediction of electrolyte performance, the design of new high-performing electrolytes, and the extraction of important scientific insights.

In the first part, we introduce a novel workflow that combines principles of feature engineering, feature selection, and machine learning model assessment. This workflow allows us to extract insights that guide the design of five new, high-performing electrolytes. Leveraging simple features, such as elemental composition that encodes pertinent physics within the electrolytes, we constructed interpretable models using linear regression, random forest, and bagging techniques. Through the results derived from these interpretable models, we identified crucial electrolyte features that are instrumental in achieving high battery efficiency. One such feature is the atomic fraction of oxygen in the solvent, highlighting the significance of reducing solvent oxygen for achieving high Coulombic efficiency (CE). Equipped with this insight and a few others, we formulated five new electrolyte compositions with fluorine-free solvents, one of which attains a high CE of 99.70%.

In the second part, we employ data segmentation in conjunction with machine learning methods to discern crucial performance descriptors within distinct electrolyte efficiency classes. Through this approach, we made a surprising discovery. Common electrolyte performance descriptors like lithium morphology, ionic conductivity, solid electrolyte interphase chemistry, and lithium-electrolyte reactivity, do not explain performance variations in electrolytes beyond a Coulombic Efficiency (CE) of 98%. By utilizing new machine learning model assessment techniques, interpretable machine learning models, correlation analysis, and rigorous spectroscopy and electrochemistry characterizations, we unveil the pivotal role of galvanic corrosion in accounting for performance disparities within high CE (>98%) electrolytes.

This work underscores the potential of data-driven approaches in expediting the discovery of high-performance electrolytes for lithium metal batteries.