Accelerating Catalysis Understanding via Large Language Model Data Extraction and Shallow Machine Learning...

Area of Science:

• Catalysis
• Materials Science
• Artificial Intelligence

Background:

• Catalysis research is complex, with limited understanding of microenvironments and reaction mechanisms hindering AI-driven catalyst design.
• Existing scientific literature provides insufficient high-quality data for deep learning models in catalysis.
• Machine learning applications for experimental catalysis researchers remain an open question.

Purpose of the Study:

• To present a framework leveraging large language models (LLMs) for automated generation of experimental catalysis datasets from scientific literature.
• To demonstrate the utility of shallow learning models with posthoc interpretability for extracting insights from low-fidelity datasets.
• To accelerate catalyst design and hypothesis generation for experimental researchers.

Main Methods:

• Utilized large language models to extract textual data from trusted sources, creating large, low-fidelity experimental catalysis datasets.
• Employed prompt engineering, data encoding, and question architectures for effective information extraction.
• Applied shallow learning models with posthoc interpretability to analyze the generated datasets.

Main Results:

• Successfully generated large, low-fidelity datasets for model reactions like CO2 electroreduction and oxygen reduction reaction.
• Demonstrated that shallow learning models can extract meaningful information from these datasets.
• Uncovered established facts (e.g., Cu catalytic properties) and novel insights (e.g., voltage thresholds for multicarbon products).

Conclusions:

• The proposed framework enables experimental catalysis researchers to utilize machine learning for rapid literature processing.
• It facilitates the generation of novel hypotheses and accelerates the design of new catalysts.
• This approach serves as an entry point for integrating AI into experimental catalysis workflows.