Multi-Temperature VOCs + OH Rate Coefficients Prediction via GCN Constrained by the Arrhenius Equation

Based on the accumulation of the high-quality databases from experimental measurements and theoretical calculations, data-driven machine learning approaches has been widely applied to rate coefficient prediction. However, most existing models are confined to single-temperature predictions and rely primarily on data-driven representations, resulting in limited temperature extrapolation capability. Although Arrhenius parameters have been introduced in some studies to encode temperature dependence, their effective utilization is restricted by the scarcity of high-quality Arrhenius-parameter datasets. To address these challenges, this study proposes a Graph Convolutional Network (GCN) architecture that deeply integrates physical constraints with data-driven strategies, leveraging an atmospheric-chemistry database of reaction kinetics between Volatile Organic Compounds (VOCs) and OH radicals. Instead of directly predicting rate coefficients, the model predicts Arrhenius parameters and indirectly derives rate coefficients across multiple temperatures, thereby enforcing physics-informed regularization over a broad temperature range. Meanwhile, to overcome the scarcity of high-quality Arrhenius parameters, we employed geometric data augmentation and a self-training-based semi-supervised learning strategy. This approach successfully incorporated 821 semi-labeled samples (containing only rate coefficient data) from the dataset, maximizing the utility of available kinetic data. Evaluation results demonstrate that the proposed model achieves excellent predictive performance (R2 = 0.93), significantly outperforming baseline GCN models (R2 = 0.77), with ablation studies verifying the vital contribution of each module. Functional group analysis confirms strong generalization for H-abstraction, while applicability domain analysis shows broad utility except for polyhalogen compounds. This work provides a novel methodological framework merging physical consistency and data-driven learning, which enables rapid, wide-temperature kinetic predictions under conditions of small and incomplete datasets. Learning from this database facilitates its convenient application to atmospheric pollutant screening and environmental risk assessment.