Machine learning modeling in flow-electrode capacitive deionization system: Prediction of ion concentrations in flow-electrode aqueous electrolytes

Abstract

Among various desalination technologies, flow-electrode capacitive deionization (FCDI) has emerged as a promising electro-driven method for addressing water scarcity. This study focuses on developing simple and efficient models for predicting ion adsorption concentrations in FCDI systems. We constructed four different machine learning (ML) models using data from previously reported short-circuited closed-cycle FCDI systems. The data covered a wide range of ion concentrations in flow-electrode electrolytes, revealing asymmetric electroadsorption behavior for Na+ + and Cl- ions. This difference can be explained by the difference in the abundance of electron acceptors in the cathode chamber. Notably, the nearest neighbors and gradient boosting algorithms demonstrated high predictive performance (R2 2 > 0.99) for sodium ions (MAE: 448.3-457.3) and chloride ions (MAE: 422.2-457.3) with computation times under 80 s. Additionally, voltage and biochar weight percentage in the flow-electrode were identified as key factors in ML-based FCDI modeling. This study lays the groundwork for integrating ML into FCDI systems for sustainable desalination.