Elucidating per- and polyfluoroalkyl substances (PFASs) soil-water partitioning behavior through explainable machine...

Area of Science:

• Environmental Chemistry
• Soil Science
• Computational Chemistry

Background:

• Per- and polyfluoroalkyl substances (PFASs) are persistent environmental contaminants.
• Understanding PFASs soil-water partitioning is crucial for assessing their environmental fate and migration.
• Predictive models can enhance the understanding of PFASs behavior in complex soil environments.

Purpose of the Study:

• To develop and optimize a machine learning model for predicting PFASs soil-water partitioning coefficients (Kd).
• To identify key soil properties and PFAS characteristics influencing soil-water partitioning.
• To assess the potential for environmental migration of PFASs based on predicted partitioning.

Main Methods:

• An optimized random forest (RF) model was developed to predict Kd values.
• The model utilized 11 easily obtainable features, including molecular weight and soil pH.
• Three-dimensional interaction analyses were performed to identify specific partitioning conditions.

Main Results:

• The RF model achieved high predictive performance (R² = 0.93, RMSE = 0.86).
• High organic carbon (OC), cation exchange capacity (CEC), and low soil pH correlated with higher Kd values (stronger adsorption).
• Low Kd values, indicating weaker adsorption and higher migration potential, were observed in high pH, low OC, and low CEC soils with lighter molecular weight PFASs.

Conclusions:

• Machine learning models, like RF, are powerful tools for understanding PFASs partitioning behavior.
• Soil properties (OC, CEC, pH) and PFAS molecular weight are critical determinants of soil-water partitioning.
• The study provides valuable insights for predicting PFASs environmental distribution and migration.