Carbon-Supported High-Entropy Electrocatalyst Design via Large Language Models

Abstract Carbon-supported high-entropy alloy (HEA) electrocatalysts are promising for electrochemical reactions, but rational design is hindered by an enormous composition and structure search space. Reusing prior knowledge can accelerate rational design because it provides constraints and trends that shrink the effective search space. It helps prioritize compositions and structures with higher likelihood of activity and stability, and it reduces repeated testing of known failure modes. In practice, this still poses a challenge as the knowledge is fragmented across a large and time-spanning literature. Large language models (LLMs) can read relevant literature at scale, but turning unstructured text into design guidance for electrocatalysts is still a key gap. Here, we develop CarbonCat-LLMs, a framework that uses LLM-based extraction from a publication’s abstract and cosine-similarity analysis of learned embeddings. It prioritizes compatible multi-element HEA compositions and carbon supports for hydrogen evolution reaction (HER) across a search space of millions of candidates. As a proof of concept, CarbonCat-LLMs achieves state-of-the-art performance on benchmark carbon-supported HER electrocatalyst design tasks. In time-split tests, it identifies high-performing catalyst configurations years before they appear in later experimental reports. By leveraging decades of published literature with LLM-based extraction and analysis, this work provides a scalable, knowledge-driven route to accelerate electrocatalyst discovery and design.