This research paper presents a novel approach to scientific reasoning in catalyst design. The approach uses a Monte Carlo Tree Search-based method in conjunction with large language models to perform a goal-driven combinatorial search. The method improves upon state-of-the-art chain-of-thought prompting variants to enhance scientific reasoning. The study introduces two new reasoning datasets related to computational chemistry simulations and questions posed by catalysis researchers. The proposed approach outperforms the best baseline by 25.8% and provides novel insights that can enhance the reasoning and discovery process for scientists.
Publication date: 22 Oct 2023
Project Page: https://github.com/pnnl/chemreasoner
Paper: https://arxiv.org/pdf/2310.14420
