Machine Learning Prediction and Model Interpretation of Nitroxide–Radical Recombination Kinetics for Nitroxide-Mediated Polymerization

Abstract Nitroxide-mediated polymerization (NMP) is a versatile controlled radical polymerization technique compatible with industrial processing conditions. The level of control in NMP depends strongly on the structure of the nitroxide mediator, whose design remains largely empirical due to the limited understanding of structure–reactivity relationships. A key kinetic parameter governing this process is the nitroxide–radical recombination rate coefficient, kc. To address this, we developed an interpretable machine-learning model to predict log10kc from molecular descriptors of the three reaction components: the nitroxide, the carbon-centered radical, and the solvent. The model was trained on a dataset of 305 experimental recombination rate coefficients. Molecular representation was optimized within a nested cross-validation workflow. The best performance was achieved with an XGBoost model that combined RDKit and ECFP descriptors, yielding R2 = 0.83 and RMSE = 0.29 for log10kc. SHAP (SHapley Additive exPlanations) analysis was used to interpret the model and showed that it captures the main factors controlling recombination kinetics. The model also reproduces expected trends: bulkier, more lipophilic nitroxides and stabilized aromatic radicals recombine more slowly. Component-level attribution showed that the radical contributes 51% of the predictive signal, followed by the nitroxide (34%) and the solvent (15%).