MOF-Sleuth Audits Chemical Files with Explainable AI

This research introduces MOF-Sleuth, an AI agent designed to audit crystallographic information files (CIFs) for Metal-Organic Frameworks (MOFs), which are crucial for computational chemistry. MOF-Sleuth addresses the limitations of existing methods that either lack fine-grained explanations or struggle with reliable chemical reasoning. The agent comprises two main modules: a deterministic Forensic Lab that extracts detailed chemical evidence (composition, geometry, connectivity, etc.) and a Sleuth reasoning engine that uses this evidence to generate explanations, identify error types, and make a binary decision. The system employs reward-guided reinforcement learning to align tool measurements with chemical explanation-level supervision, ensuring that not only the final answer but also the cited evidence and supported diagnoses are rewarded. A new metric, Chemically Grounded Diagnosis (Chem-GD), assesses the quality of these explanations. MOF-Sleuth achieves state-of-the-art performance across benchmarks, demonstrating significant improvements in error detection, attribution, and the quality of grounded explanations compared to other LLM-based and MOF-specific machine learning approaches.