Interpretable-generative machine learning approaches for predicting simultaneous removal of organic pollutants and heavy metals from water by adsorbent materials

Abstract

This study constructed a dataset containing commonly used adsorbents for the simultaneous removal of organic pollutants (OPs) and heavy metals (HMs) from aqueous phases, and evaluated machine learning models to predict the interactions between OPs and HMs on the adsorbents. Data synthesis using Wasserstein generative adversarial networks (WGAN) and feature importance interpretation via Shapley additive methods were analyzed. The results revealed that categorical boosting (CatBoost) emerged as the best machine learning model due to its excellent adsorption prediction performance and model interpretability. The interactive effects were classified into “inhibition”, “no significant effect” and “promotion” categories. The original dataset was computationally augmented with 1500 reliable synthetic data by WGAN, improving the precision for “promotion” category of CatBoost from 0.79 to 0.93. Adsorbent surface area and solution pH were the primary factors causing the “promotion” of adsorptive behavior. Bridging and electrostatic interaction mechanisms may play important roles in promoting simultaneous adsorption. The strategy of combining interpretable machine learning and data synthesis with the WGAN method, not only provides new insights into the prediction of the composite adsorption behavior of OPs and HMs by adsorbents in the aqueous phase, but also addresses the problem of insufficient experimental data to improve the prediction performance.