Large Language Model-Assisted Development of Reliable Machine Learning Models for Hydrogen Evolution Reaction Catalysts

Abstract Reliable experimental datasets are essential for accelerating sustainable hydrogen production through data-driven electrocatalyst discovery. Large language models (LLMs) have enabled the rapid construction of such datasets by extracting catalytic features directly from unstructured scientific literature. However, the structural reliability of such datasets for machine learning (ML) applications remains poorly understood. In this work, the reliability of LLM-extracted datasets for hydrogen evolution reaction (HER) catalysts is systematically evaluated. It is shown that high feature-level extraction accuracy does not guarantee instance-level correctness when multiple catalyst features must be jointly valid, leading to multiplicative error propagation across data records. To address this challenge, a quality control (QC)-aware data construction framework is developed. When evaluated under an identical ML pipeline, the QC-refined dataset recovered approximately 83% of the achievable predictive performance improvement toward the manually curated golden dataset. The resulting QC-refined dataset further enables compositional performance mapping of HER catalysts, revealing activity trends across dopant–active phase combinations and providing insights that can help guide HER catalyst design.