Machine learning-based prediction of PAHs thermal desorption efficiency: Model optimization, boundary correction, and mechanistic insights

Pengcheng Fu, Meng Qi, Chunshuang Liu, Cheng Zhang, Yuhao Zhao, Zhibao Liu, Hao Wang, Dongfeng Zhao,
Machine learning-based prediction of PAHs thermal desorption efficiency: Model optimization, boundary correction, and mechanistic insights,
Journal of Environmental Management,
Volume 395,
2025,
127705,
ISSN 0301-4797,
https://doi.org/10.1016/j.jenvman.2025.127705.
(https://www.sciencedirect.com/science/article/pii/S0301479725036813)
Abstract: Accurate prediction of PAHs removal efficiency during thermal desorption remains challenging due to complex nonlinear interactions among soil properties, contaminant characteristics, and operational parameters. To address this, nine machine learning models were developed and optimized to predict PAHs thermal desorption efficiency, incorporating Bayesian hyperparameter tuning, boundary-aware target transformation, and interpretability analysis. A logit-based transformation was applied to ensure the physical plausibility of remediation efficiency [0–100 %] and improve prediction accuracy near the upper boundary (>90 %). Among the tested models, ensemble tree algorithms, particularly the CatBoost model (R2 = 0.9709), achieved the highest accuracy without overfitting. Shapley Additive Explanations and partial dependence analyses revealed that reaction time, temperature, and the boiling point-to-temperature ratio were dominant predictors, while soil moisture and organic matter exhibited dual inhibitory–facilitative roles, confirming two-phase desorption kinetics and threshold-based thermal effects. The optimized models maintained strong generalization in 1200 Monte Carlo validations (R2 > 0.94), and stratified modeling by pollutant species further improved accuracy (R2 > 0.98 for phenanthrene and pyrene). This study establishes a robust and interpretable machine learning framework for optimizing PAHs thermal desorption and provides practical insights for enhancing remediation reliability, energy efficiency, and intelligent process control.
Keywords: Polycyclic aromatic hydrocarbons; Thermal desorption; Machine learning; Model optimization; Target transformation