Optimal Management Strategy for Salt Adsorption Capacity in Machine Learning-Based Flow-Electrode Capacitive Deionization Process

Flow-electrode capacitive deionization (FCDI) has created a breakthrough toward a more stable desalination performance by adopting a flow-electrode compared to existing capacitive deionization and membrane capacitive deionization as a promising electrochemical water treatment technology. However, the FCDI technology requires investigation of various mechanisms pertaining to flow-electrode materials to achieve system optimization. Further, studies on applying machine learning to the FCDI technology have been scarcely reported. Our study aims to explore optimal algorithms via machine learning for predicting the salt adsorption capacity of FCDI processes and evaluate the feasibility of optimization applications. Concurrently, a comparative analysis was conducted through the performance model indicators of mean absolute error (MAE), mean squared error, and R² for support vector machine, random forest, and artificial neural network (ANN) algorithms. Herein, we demonstrated that the optimal ANN-based model exhibited the highest predictive performance, achieving R² and MAE values of 0.996 and 0.21 mg/g, respectively. Additionally, the Shapley additive explanations (SHAP) confirmed a trend in the contribution of influent concentration, aligning closely with the results of statistical analysis. Specifically, the change in voltage of the FCDI process serves as a key factor in determining salt adsorption efficiency. Moreover, a parallel comparison of the Pearson correlation coefficient and SHAP analyses suggests that the impact of voltage entails a nonlinear contribution within the realm of machine learning. Finally, to deploy a machine learning-driven ANN model system, we present multiple factors (e.g., weight of flow-electrodes, influent concentration, and voltages) as a reinforcement learning model for decision-making. This offers valuable insights and guidance for future operations of the FCDI process.