Data-Driven Interrogation of Reactivity in Acid-Catalyzed Carbonyl-Olefin Metathesis with Machine Learning and Large...

Area of Science:

• Organic Chemistry
• Catalysis
• Computational Chemistry

Background:

• Carbonyl-olefin metathesis (COM) is a key reaction for C-C bond formation.
• Predicting catalyst performance in COM is challenging due to mechanistic complexity.
• Machine learning (ML) and large language models (LLMs) offer potential for experimental design.

Purpose of the Study:

• Develop a practical ML framework to guide catalyst and substrate selection for COM.
• Address the challenges in predicting COM reactivity and catalyst effectiveness.
• Compare the strengths of ML and LLMs in mechanistically complex catalytic systems.

Main Methods:

• Compiled a curated, blinded dataset of 147 COM reactions.
• Developed tiered predictive models including a Morgan-fingerprint baseline and quantum-derived descriptors.
• Employed feature-importance, SHAP analyses, reactivity-cliff analysis, and Bayesian optimization.
• Utilized GPT-4o for a code-free, LLM-guided experimental selection protocol.

Main Results:

• Quantum-derived descriptor models achieved high accuracy (R² ≈ 0.92) and external validation.
• Identified key reactivity drivers: catalyst HOMO energy, dimerization propensity, and substrate carbonyl polarization.
• Reactivity-cliff analysis distinguished predictable from unpredictable reaction systems.
• Variance-guided Bayesian optimization and LLM-guided selection significantly reduced experimental effort.

Conclusions:

• Established a mechanistically interpretable, data-efficient framework for COM experiment prioritization.
• Demonstrated the utility of uncertainty-aware ML and LLM-guided selection for optimizing catalytic reactions.
• Provided a practical approach to augment decision-making in complex catalytic systems like COM.