Abstract Designing electrolyte materials for high-energy lithium metal batteries requires navigating vast, discrete chemical spaces, where intricate interphasial and electrolyte chemistries render component interactions largely unclear. Traditional trial-and-error methods struggle with discontinuous electrolyte-performance relationships and inefficient adaptation to new molecular candidates, hindering discovery. Here, we propose a two-stage deep active learning framework with knowledge transfer for rapid electrolyte design. In stage one, deep active learning with deep kernel learning selects informative experiments and models discontinuous relationships between formulation and performance, improving sample efficiency and reducing experimental cost. In stage two, target statistic coding quantifies what was learned and transfers it to new design settings, such as expanded formulation spaces and newly introduced components, using only a small number of additional measurements. Using this framework, we found electrolytes that increase the average lifetime of lithium metal symmetric cells by threefold after three learning iterations, and we rapidly identified improved formulations for Li 0 | |LiNi 0.8 Co 0.1 Mn 0.1 O 2 full cells in expanded chemical spaces. This work provides an experiment-driven, sample-efficient route to explore complex electrolyte formulation spaces and quantify inter-component correlations, as well as a realistic, high-cost, small-data benchmark for probabilistic surrogate modeling and sequential decision-making in discrete chemical spaces.
Hong et al. (2026) studied this question.