Augmented XGBoost Modeling with CNN-extracted Mass-spectral Features and Temporal Embedding for Enhanced Prediction of Pyrene Degradation Kinetics in Advanced Oxidation Processes

Advanced Oxidation Processes (AOPs) are widely employed for the degradation of polycyclic aromatic hydrocarbons (PAHs). Determining the optimal oxidation method and corresponding process parameters is crucial for achieving the maximum degradation efficiency. Due to the complexity of AOPs, conventional kinetic models and machine learning approaches show limited accuracy in predicting the degradation kinetics of PAHs, making them inadequate in optimizing process parameters for engineering practices. This study pioneers the use of convolutional neural networks (CNN) for deep feature extraction from the gas chromatography-mass spectrometry (GC-MS) data of pyrene degradation. Meanwhile, the Fourier-based temporal embedding (FTE) method was adopted to map the discrete time into continuous frequency-domain features, preserving a temporal continuity across reaction trajectories. The fusion of CNN-extracted MS features, FTE, and reaction conditions within an eXtreme Gradient Boosting (XGBoost) framework yielded a novel CNN-FTE-XGBoost hybrid model, enabling a cross-scale modeling of degradation kinetics. The CNN-FTE-XGBoost architecture demonstrated exceptional prediction accuracy on pyrene degradation efficiency ( R ²=0.951, RMSE=0.046, MAE=0.039). SHapley Additive exPlanations (SHAP) analysis further identified the key factors influencing pyrene degradation efficiency, verifying the interpretability of model. This work establishes an interpretable multimodal fusion-driven modeling framework for a high-precision prediction of pollutant degradation in AOPs, providing a reliable tool for AI-guided optimization of AOPs in remedial applications.