Optimizing capacitive deionization operation using dynamic modeling and reinforcement learning

Abstract

Capacitive deionization (CDI) is a promising desalination technology through an electrochemical mechanism, especially for brackish water. The energy efficiency of the CDI process depends on the complex relationship among several factors, such as flow rate, voltage or current, charging/discharging times, and the water quality of the influent. To minimize the energy consumption of the CDI system, we used deep reinforcement learning (DRL), an effective optimization technique. The COMSOL-CDI model, a powerful tool for computational fluid dynamics, was used as the DRL environment to address the limitations of one-dimensional numerical CDI models in accurately predicting complex ion transport mechanisms and kinetics. Among four model-free algorithms, soft actor-critic (SAC) was selected as the best-performing model, reducing the specific energy consumption by 77.18 %, and increasing the desalted water production by 15 % under the constraint-satisfying conditions. Notably, the SAC model exhibited robust performance in testing scenarios with random influent concentrations. The automated system developed in this study can be utilized to effectively control the desalination process based on a comprehensive two-dimensional ion transport model integrated with fluid dynamics analyses.