Predicting the Partitioning Behavior of Per- and Poly-Alkyl Substances (PFAS) on Liquid-Solid Interface for Carbon and Mineral Based Surfaces using Multivariate Linear Regression Models with K-Fold Cross Validation.

Computational data science , especially the machine learning approach has been a major contribution to the field of engineering. In this study, data mining and machine learning were practiced estimating the partitioning of Per- and Poly-fluoroalkyl Substances (PFAS) compounds during aqueous adsorption on various adsorbent materials with a vision to potentially replace the time-consuming and labor-intensive adsorption experiments. Regression models, such as linear, tree, support vector machine (SVM), ensemble of trees, and gaussian process regression (GPR) models were trained and tested using previously published data. 290 data points and 170 data points for activated carbon and mineral adsorbents, respectively, were mined for training the models and 10 data points were used to test the trained models. Statistical parameters, such as Root-Mean-Square Error (RSME), R-Squared, Mean Average Error (MAE), Mean Squared Error (MSE), etc., were used to compare the regression models. It was found that rational quadratic GPR (R-squared = 0.9966) and fine regression tree (R-Squared = 0.9427) models had the highest estimation accuracy for carbon-based and mineral-based adsorbents, respectively. These models were then validated for prediction accuracy using 10 data points from previous studies as an outer test set. Rational quadratic GPR was able to achieve 99% prediction accuracy for carbon-based adsorbent, while fine tree regression model was able to achieve 94% prediction accuracy. Despite such high estimation accuracy, the data mining process revealed the data shortage and the need for more research on PFAS adsorption to present real-world models. This study, as one of the first, shed a light on the determination of key parameters in aquatic chemistry with data mining and machine learning approaches.