Predicting sorption of organic pollutants on soils with interpretable machine learning

Area of Science:

• Environmental Chemistry
• Soil Science
• Computational Chemistry

Background:

• Sorption of organic pollutants (OPs) on soils is crucial for environmental fate and transport.
• Understanding nonlinear relationships between adsorption capacity and influencing factors is limited.
• Predicting OPs sorption requires advanced modeling due to complex interactions.

Purpose of the Study:

• To develop and compare five machine learning (ML) models for predicting OPs sorption on soils.
• To identify key factors influencing OPs adsorption using interpretability analysis.
• To map OPs sorption capacities across mainland China and assess environmental risks.

Main Methods:

• Utilized a dataset of 352 data points from previous studies.
• Developed and evaluated Support Vector Machine (SVM), Deep Neural Networks (DNN), Extreme Gradient Boosting (XGBT), Random Forest (RF), and Gradient Boosting Decision Tree (GBDT) models.
• Applied Shapley Additive Interpretation (SHAP) for model interpretability and generated a spatial distribution map of OPs sorption capacities.

Main Results:

• The XGBT model achieved superior performance with R² of 0.952 and RMSE of 0.103.
• Electronic effects of OPs and soil organic matter (SOM) content were identified as the most influential factors.
• High sorption capacities were predominantly found in southern and southwestern China, correlating with reduced environmental risks.

Conclusions:

• Developed a novel interpretable ML framework for predicting OPs adsorption potential.
• Highlighted the dominant roles of π-π interactions and hydrophobic partitioning in OPs sorption mechanisms.
• The framework supports environmental management, risk assessment, land remediation, and soil protection policies.