Interpretable machine learning-enabled atomic reactivity decoding for precision design and reverse synthesis of sustainable rejuvenators

Accurate quantification and modulation of molecular reactivity is critical for the targeted design of high-performance rejuvenators, while conventional empirical and computational methods lack interpretability and applicability and impede precise rejuvenator development. Here, we propose an interpretable automated machine learning (ML) framework tailored for small datasets to enable robust classification and quantitative prediction of rejuvenator molecular reactivity, overcoming the data scarcity bottleneck in material research. The XGBoost model achieves 92.12% accuracy in classifying three types of rejuvenators. Combined SHapley Additive exPlanations and quantitative structure-property relationship (QSPR) modelling reveals the dominant regulatory role of the HOMO-LUMO gap. Four ML-QSPR models (TabICLv2, TabPFN, MLP, and XGBoost) realize reliable HOMO-LUMO gap prediction with coefficients of determination over 0.82 for the test set. Further SHAP analysis identifies core descriptors and their nonlinear threshold effects. Notably, we develop an original atomic reactivity decoding heatmap embedded in our small-data adaptive framework for intuitive atomic-scale reactivity visualization for rejuvenators, validated by conceptual density functional theory calculations and experimental tests. This work integrates small-data-driven property prediction, innovative atomic-level reactivity visualization, and interpretable mechanism analysis, providing a generalizable strategy for rational design and rapid screening of functional rejuvenators and organic materials.