@

초록

Overcoming the polysulfide shuttle effect is the critical challenge for developing high-performance potassium-sulfur (K-S) batteries. While designing catalyst materials is a promising route, the vast compositional space makes traditional screening impractical. Machine learning (ML) offers an approach to accelerate this process but suffers from the challenge of creating comprehensive feature representations and a lack of interpretability. To address these two challenges, we propose a tri-modal machine learning framework that interactively integrates graph, text, and elemental information. This framework employs a sophisticated fusion strategy based on multi-head cross-attention, which creates a rich, context-aware, and synergistic embedding that serves as a powerful feature representation. Moreover, it uses a component-level approach to interpretability by validating the quality of the textual features through attention score analysis. Using this framework, a high-throughput screening of over 75,000 configurations identified Ni-doped and Zn/Ti-co-doped CuS as promising candidates. Their superior electrochemical performance was confirmed by experiments. Further mechanistic insights from in situ and ex situ characterizations corroborate accelerated sulfur conversion kinetics. Additionally, the framework’s versatility was validated on diverse external datasets, spanning chemically and structurally distinct material systems and their associated bulk and surface properties. This work establishes a validated tri-modal ML framework that can provide a rapid, reliable, and interpretable pathway to accelerate the discovery of diverse, high-performance energy materials.

키워드