Mechanism-informed hierarchical machine learning for predicting kinetic isotope effects in •OH-mediated reactions.

Deuterated compounds are increasingly applied in pharmaceuticals and agrochemicals, yet predicting their environmental transformation kinetics remains challenging because isotope-dependent kinetic data are scarce and isotope effects represent subtle perturbations superimposed on intrinsic molecular reactivity. Here, we develop a hierarchical Δ-learning framework that separates molecular reactivity into baseline reactivity and isotope-induced kinetic perturbations, enabling data-efficient prediction of kinetic isotope effects (KIEs) under limited-data conditions. The framework integrates low- and high-fidelity kinetic information through three sequential stages: learning baseline reactivity from experimental rate constants of non-deuterated compounds (kH), capturing isotope-induced perturbations from computational Δlog(k) data, and refining predictions using limited high-fidelity quantum chemical kinetic data. The framework achieves reliable predictive performance (cross-validation R2 = 0.81) and reveals pathway-associated isotope sensitivity trends related to hydrogen atom transfer (HAT) and radical adduct formation (RAF) reactions. Deuterated rate constants (kD) can be estimated by combining predicted isotope-induced perturbations with available kH values. Applications to representative environmental contaminants demonstrate that isotopic substitution can alter apparent transformation kinetics during •OH-mediated oxidation, with the magnitude of isotope effects dependent on molecular structures and reaction characteristics. Beyond KIE prediction, this work highlights a general multi-fidelity learning strategy for extracting weak mechanistic signals from limited high-quality data and provides a foundation for data-efficient screening of emerging deuterated contaminants.