Bayesian Optimization of Catalysis with In-Context Learning

Area of Science:

• Artificial Intelligence
• Materials Science
• Chemistry

Background:

• Large language models (LLMs) excel at classification via in-context learning (ICL), adapting to new tasks without weight updates.
• Current materials discovery often involves extensive characterization of suboptimal materials, slowing innovation.

Purpose of the Study:

• To extend in-context learning (ICL) capabilities of frozen LLMs to regression tasks with uncertainty estimation.
• To enable Bayesian optimization (BO) for materials discovery using natural language prompts, bypassing traditional training and feature engineering.

Main Methods:

• Representing materials as synthesis and testing procedures within natural language prompts for LLMs.
• Applying Bayesian optimization with in-context learning (BO-ICL) for a design-first materials discovery approach.
• Utilizing frozen LLMs (e.g., GPT-4o, Gemini) for regression and uncertainty estimation.

Main Results:

• BO-ICL demonstrated performance matching or exceeding Gaussian processes on aqueous solubility and oxidative coupling of methane (OCM) benchmarks.
• In reverse water-gas shift (RWGS) experiments, BO-ICL rapidly identified high-performing multimetallic catalysts from large candidate pools.
• Achieved near-equilibrium CO yield within 6 and 10 iterations for datasets of 3,700 and 360,000 materials, respectively.

Conclusions:

• The novel BO-ICL method redefines materials representation and significantly accelerates discovery processes.
• This approach offers broad applicability in catalysis, materials science, and artificial intelligence, paving the way for faster innovation.