Application of machine learning in microwave remediation of total petroleum hydrocarbon contaminated soil: Prediction and key factor identification.

Microwave thermal remediation (TPH) is a promising remediation method for petroleum hydrocarbon contaminated soils due to its high energy efficiency and rapid heating capacity. However, the complexity of influencing factors and their nonlinear interactions often hinder the quantitative understanding of TPH removal rates. In this study, a prediction dataset containing 217 instances was constructed. The researchers developed four machine learning models Random Forest (RF), Support Vector Regression (SVR), Backpropagation Neural Network (BPNN), and Extreme Gradient Boosting (XGB) to predict TPH removal rates. Among them, the XGB model performed better (RMSE = 8.22 %, R = 0.8832). In order to improve the generalization and computational efficiency of the model, a Bayesian optimization algorithm based on tree-structured Parzen estimation (TPE) was used, and an optimized XGB model was obtained, with the RMSE reduced to 4.58 % and the R improved to 0.9658. Interpretability analysis using SHAP and PDP showed that the microwave temperature was the most important influencing factor, followed by the treatment time and the water content. In addition, the prediction uncertainty was quantified using bootstrap resampling. The proposed modeling framework provides not only accurate predictions but also insights into key operational parameters, with potential applications in adaptive soil remediation planning and intelligent decision support systems.