Extracting Electrolyte Design from Interpretable Data-Driven Methods

Data-driven machine learning (ML) models have demonstrated tremendous success in learning the complexity of material design, enabling them to make reliable predictions of associated physical and chemical properties. However, a good predictive model alone falls short in drawing forth an interpretation of the correlation between target property and the changing material design. Therefore, an interpretable framework must accompany the predictive model to elaborate the understanding developed by the model and extract a set of rules for material design. In this talk, we discuss state-of-the-art data-driven approaches to map electrolyte formulation design to the performance of the battery electrolyte. This includes dense feed-forward networks, graph-based models, and transformer-based foundation models. Next, we extract design rules from these models using an interpretability framework to develop electrolytes with target performance. The approach is successfully applied to design electrolytes custom to cathode loadings in a novel multi-electron reaction-driven interhalogen battery based on I-Cl chemistry. It is a common observation in conversion batteries that increments in the weight of active cathode material (cathode loading) after a critical point can significantly drop battery’s performance due to internal resistance, shuttling, and side reactions. However, the strong codependence of electrolyte formulation design on cathode loading can be leveraged to achieve higher performance in a battery by a compatible electrolyte design. In the present study, the electrolyte designs suggested by the data-driven approach demonstrate an additional 20% improvement in the interhalogen battery capacity over experimental optimization at targeted cathode loading. The study evaluates the suitable methods to leverage artificial intelligence (AI) for electrolyte discovery by considering chemical information of constituent materials, their composition, and additional device design variables such as separators and active material loadings.