Hierarchical Moiety-Aware Graph Transformer for Li-metal Electrolyte Formulation Design

Maximizing Coulombic efficiency (CE) is critical in improving battery life and performance, although existing predictive models overlook the very chemical design strategies that drive electrolyte performance. Here, we introduce HELENA (Hierarchical Electrolyte Learning with Embedded Neural Attention), the first electrolyte deep‑learning framework to encode moiety‑level chemistry and cooperative interactions across full formulations. HELENA constructs a three‑tiered representation and employs attention mechanisms to learn both established heuristics and potentially unseen design principles. Trained on a substantially expanded CE dataset of lithium metal batteries, HELENA outperforms state‑of‑the‑art graph‑ and tree‑based models, while its attention mechanisms both recover established electrolyte trends and reveal novel moiety co‑dependencies. Applying HELENA as an in-silico screening engine, we experimentally validate several model‑predicted formulations achieving CE > 99.5%, demonstrating its practical utility. By fusing moiety-aware tokenizer with scalable attention‑based learning, HELENA establishes a new paradigm for rational electrolyte discovery and paves the way for generative, closed‑loop design workflows of exploring multicomponent chemical systems beyond energy.