Leveraging Machine Learning Models To Optimize Electrochemical CO ₂ Capture from Treated Used Water

Dissolved inorganic carbon can be captured from treated water using electrochemical systems, thereby diminishing greenhouse gas emissions from the water sector. However, large spatiotemporal variability in the treated water chemistry can affect the efficiency of the electrochemical step. In this study, two predictive models, Artificial Neural Networks (ANN) and Random Forests (RF), were used to simulate an electrochemical CO2 capture system and rank the importance of water chemistry features and operational parameters on the efficiency of CO2 removal from treated water. After data preprocessing and hyperparameter optimization, the models were trained on a total of 252,000 data points. Among the two predictive models used, the RF model demonstrated superior performance in terms of training efficiency (<349 s for the training with hyperparameter optimization), computational cost (<1.1 s for simulating with the trained model), and predictive accuracy (R2 > 0.998 and NRMSE < 2.26). Interpretation of the trained models enabled quantification of the relative impacts of influent characteristics and operating conditions on the CO2 capture performance, indicating that operational parameters can be adjusted to compensate for water chemistry variations. The RF model was then used to determine the optimal operating conditions during electrochemical CO2 capture with variable influent composition. Dynamic adaptation of operational parameters following model optimization resulted in a 48% improvement in energy efficiency, along with an increase in overall CO2 capture efficiency. Collectively, these results indicate that machine learning models can be used to account for the spatiotemporal variability of water chemistry in treated effluents and ensure stable and efficient CO2 removal.