Prediction of Propene Hydroformylation using Machine Learning

Here, we develop machine learning (ML) models to predict catalytic performance in rhodium-catalyzed propene hydroformylation and identify new ligands and reaction conditions for enhanced yield and selectivity. A curated dataset of 215 experimentally obtained data points involving [Rh(acac)(CO)₂] precatalyst and 49 phosphine ligands was used. Molecular, electronic, and reaction descriptors were constructed using cheminformatic representations and dimensionality reduction. Among several algorithms tested, XGBoost exhibited the best performance, achieving a root-mean-square error (RMSE) of 9.85% for iso-selectivity under leave-one-ligand-out cross-validation. Feature importance analysis revealed that ligand and solvent descriptors most strongly influence selectivity, whereas turnover number (TON) predictions were more challenging. A bootstrapped ensemble of XGBoost models integrated with a genetic algorithm enabled the exploration of vast ligand–condition space, yielding 58 candidate ligands predicted to achieve TON ≥ 330 and Iso(%) ≥ 70. This study demonstrates that interpretable ML models can complement mechanistic understanding and accelerate catalyst design for small-alkene hydroformylation.