ARCHE Enables Autonomous Chemical Mechanism Discovery with Agentic Reasoning

Summary

The paper introduces ARCHE, an autonomous agentic system designed to automate chemical reaction-mechanism investigations, which typically require substantial expert intervention. ARCHE combines a general-purpose reasoning model, a domain-specialized computational chemistry model, and a structured tool registry. It interprets scientific questions, proposes and prioritizes mechanistic hypotheses, orchestrates computational workflows, and revises its conclusions in a closed loop based on computed evidence. The authors evaluate it in three progressively demanding settings: reconstructing stereocontrolling transition states and validating a previously reported asymmetric catalytic mechanism; proposing and validating a plausible radical pathway for a recently discovered but unpublished alpha-iodoboronate C-I cleavage reaction; and identifying a chemically interpretable selectivity descriptor for nickel-catalyzed migratory cross-coupling. The results are presented as evidence that combining agentic reasoning with computational validation can support autonomous mechanistic discovery. The authors describe ARCHE as a foundation for broader machine-assisted chemical research, and make its code publicly available through the linked GitHub repository.