Large language model-guided high-throughput screening for accelerated discovery of Pt-based high-entropy electrocatalysts.

Abstract

Pt-based high-entropy alloy (Pt-HEA) catalysts show significant potential for the hydrogen evolution reaction (HER), yet their vast compositional space challenges conventional trial-and-error approaches. Here, we propose a large language model (LLM)-guided closed-loop framework for high-throughput discovery to accelerate the rational design and optimization of Pt-HEA catalysts. The framework first extracts structured knowledge from extensive literature and incorporates retrieval-augmented generation (RAG) to develop an LLM assistant tailored to HEA electrocatalysis research. The assistant transforms scattered literature knowledge, experimental constraints, and platform capabilities into experimentally feasible Pt-HEA candidate combinations and executable experimental protocols. High-throughput synthesis and screening enabled rapid evaluation of multiple Pt-HEA candidates, followed by Bayesian optimization to efficiently refine compositions within the representative Fe-Co-Ni-Cu-Pt system. This established a collaborative closed-loop workflow linking LLM-assisted candidate generation, high-throughput feedback, Bayesian optimization recommendations, density functional theory (DFT) mechanistic interpretation, and proton exchange membrane water electrolysis (PEMWE) device validation. Using this framework, we completed three iterative cycles involving 50 samples within 48 h, revealing how HER activity depends on atomic ratios in the Fe-Co-Ni-Cu-Pt system and identifying Fe₃₀Co₃₀Ni₁₀Cu₁₀Pt₂₀ as the optimal composition. Theoretical calculations demonstrated that the multimetallic local coordination environment formed by Fe, Co, Ni, and Cu regulates the electronic structure of Pt active sites and fine-tunes the Pt-H interaction. Further PEMWE testing confirmed the catalyst's promising performance under device operating conditions. This work establishes a generalizable closed-loop paradigm integrating literature knowledge, experimental feedback, and model-driven decision-making, offering a viable route to the rational design and efficient optimization of complex multicomponent electrocatalytic materials.