EAGER: Accelerating decarbonization by representing catalysts with natural language

University Of Rochester, Rochester NY

Investigators

Abstract

Artificial intelligence (AI)-directed design of experiments is poised to transform chemical catalysis. Instead of using traditional structural and electronic features of catalysts characterized by expensive and difficult experiments, the project leverages recent progress in large language models (LLMs) to represent catalysts by the text of synthesis procedures and reaction conditions. The approach has potential to accelerate discovery of earth-abundant, active, and selective catalysts to bring rise to an emerging carbo-chemical industry that makes low-cost products from carbon dioxide (CO2). The broader impacts of this project will serve two purposes: (1) Educate and excite students about LLMs and AI for materials discovery; and (2) demonstrate that language-based representations are universal and can be applied to any process that is expressed with language. The LLM methodology accelerates catalyst predictions with natural language processing and Bayesian optimization (BO), while leveraging chemical intuition to develop hypotheses that guide the size and composition of the experimental space. The project focuses initially on understanding and developing trimetallic catalysts for the reverse water-gas shift (RWGS) reaction. Trimetallic catalysts are more difficult to characterize than bimetallic catalysts, making them a good fit for an approach that does not need the catalyst structure to be predictive. By representing the catalysts with language, physical-chemical details such as those due to catalyst restructuring during the reaction, are captured instead by the experimental conditions as included in the text-based representation. Beyond the core approach of utilizing language to identify novel catalysts for RWGS reaction, the project will assess the effects of experimental artifacts and irreproducible results on the model’s performance. The language-based workflow will be integrated with existing computational methods to extract mechanistic information from the text of experimental procedures. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.