JACS Au

Abstract

Catalysis is inherently complex. The lack of precise knowledge available to experimental researchers about the microenvironment, catalytic sites, mechanisms, and changes that occur under reaction conditions has hindered the effectiveness of deep-learning artificial intelligence algorithms to predict catalyst behavior under reaction conditions. Given the type and quality of data available in the scientific literature, there are still open questions on how machine learning can be used by experimentalists working in the field of catalysis to accelerate catalyst design. Here, we present a framework that leverages large language models to extract textual data from known and trusted sources to automatically generate large, but relatively low-fidelity, experimental catalysis data sets across many research groups. We also show that instead of using deep-learning models, which require high-quality data, shallow learning models with posthoc interpretability can extract valuable information about experimental catalytic systems from these low-fidelity data sets. The innovation of this work lies not in the model development but in the prompt engineering, data encoding, and question architectures employed to extract meaningful information. We applied this framework to two different model reactions: the electrocatalytic reduction of carbon dioxide and the electrocatalytic oxygen reduction reaction. We showcase that this framework has the ability to uncover known and established facts within the catalysis community, such as the catalytic properties of Cu, as well as novel insights, including the critical role of voltages above a certain threshold in producing multicarbon products from CO2. We anticipate that this proposed framework will serve as an entryway for experimental catalytic researchers to utilize machine learning to rapidly process literature, generate novel hypotheses, and design experiments to accelerate catalyst development.

Citation 

JACS Au 2025, 5, 11, 5578–5589