Deep active learning and knowledge transfer for rapid discovery of lithium metal battery electrolytes

Designing electrolyte materials for high-energy lithium metal batteries (LMBs) requires navigating vast, discrete chemical spaces, where intricate interphasial and electrolyte chemistries render component interactions largely unclear. Traditional methods struggle with discontinuous electrolyte-performance relationships and inefficient adaptation to new molecular candidates, hindering discovery. Here, we propose a novel two-stage AI framework integrating deep active learning (DAL) with knowledge transfer for rapid electrolyte design. Stage one employs DAL enhanced with deep kernel learning to efficiently identify promising electrolytes by intelligently selecting experiments and adeptly capturing discontinuous structure-property landscapes, thereby improving sample efficiency and minimizing experimental cost. Stage two utilizes a target statistic coding (TSC) to explicitly quantify and transfer learned knowledge, enabling rapid optimization within new design scenarios—including those involving novel molecular candidates—via zero-shot or few-shot learning, thus bypassing extensive de novo experimentation. Using this AI framework, we identified electrolytes achieving a three-fold average lifespan increase in Li ⁰ |Li ⁰ cells through only three DAL iterations in the first stage. Crucially, the TSC-encoded knowledge facilitated the subsequent zero-shot identification of superior electrolytes from an expanded 5,400-candidate space. We further demonstrate validated knowledge transfer for optimizing Li ⁰ |Li ⁰ |LiNi _0.8 Co _0.1 Mn _0.1 O ₂ full cells and, significantly, adapting to dynamically updated materials spaces incorporating newly discovered molecules. This AI-driven capability generated novel materials science insights, yielding high-performance ether-based electrolytes that mitigate the challenging “carbonate/ether conflict” in LMBs. This work establishes a powerful AI-driven discovery framework for rapid exploration, knowledge learning and transfer in complex materials spaces, without relying on established theories or empirical knowledge.