Machine Learning Models for Solvent Prediction in Organic Reactions: Bridging the Gap between Theory and Practical Efficacy

The selection of an appropriate solvent is crucial to achieve environmentally friendly, selective, and high-yielding chemical reactions. While artificial intelligence tools have shown promise in predicting reaction conditions, including solvent selection, they often overlook sustainability criteria in their recommendations and lack experimental validation. In this work, we present data-driven machine learning models for solvent prediction for organic reactions. Building on these models' ability to capture chemical similarity in their learned representations, we developed a green solvent replacement methodology that adapts to evolving sustainability standards without requiring model retraining. Our models achieve a Top-3 accuracy of 85.1% in predicting missing solvents for patent-derived reactions, demonstrating excellent performance even for underrepresented solvent classes. Through uncertainty analysis, we found that model misclassifications correlated with reactions that could be carried out in multiple solvents. We validated our approach experimentally, achieving a 88% success rate for general solvent prediction and a 80% success rate when focusing on green solvent alternatives. Our work addresses a key challenge in organic synthesis and provides a practical methodology for predicting both conventional and sustainable solvents for organic synthesis.