Smart control of oxychlorine species using reinforcement learning in saline electrochemical oxidation

The performance of electrochemical oxidation (EO) for remediating wastewater enriched with highly recalcitrant organics is predominantly determined by the physicochemical attributes of the selected anode materials. Together with, the wide anodic potential window unintentionally accelerates a side reaction that rapidly oxidizes influent chloride ions to perchlorate. Therefore, to effectively facilitate EO system construction, EO simulation should be designed by reflecting complex interactions among these elements. Hence, we developed a model that utilizes machine learning (ML) and reinforcement learning (RL) to predict and control the formation of oxy-chlorine species in the EO system. Specifically, herein we explore an optimal prediction algorithm to precisely control chlorate and perchlorate production under key operating parameters (e.g., electrolyte compositions, electrode types, operating time, current density, initial pH, and input chemicals). Among the four evaluated machine learning models; multi-layer perceptron (MLP), gaussian process regression (GPR), categorical boosting (CatBoost), and TabNet; the MLP architecture achieved the highest coefficient of determination and predictive accuracy (i.e., R2 = 0.775). In addition, the results of analyzing the variable contribution using Spearman correlation analysis and shapley additive explanations (SHAP) in parallel confirm that the anode material and operating time have a nonlinear effects on the formation of toxic byproducts. Ultimately, by utilizing reinforcement learning via soft actor critic (SAC) trained on data, this study develops an optimal operational strategy that boosts energy efficiency and, mitigates toxic byproduct formation.