Synergizing Data-Driven and Knowledge-Based Hybrid Models for Ionic Separations

Abstract

A hybrid modeling framework has been developed for electrodialysis (ED) and resin-wafer electrodeionization (EDI) in brackish water desalination, integrating compositional modeling with machine learning techniques. Initially, a physics-based compositional model is utilized to characterize the behavior of the unit. Synthetic data are then generated to train a machine learning-based surrogate model capable of handling multiple outputs. This model is further refined using a limited set of experimental data. The effectiveness of this approach is demonstrated by its ability to accurately predict experimental results, indicating an acceptable representation of the system’s behavior. Through an analysis of feature importance facilitated by the machine learning model, a nuanced understanding of the interaction between the chosen ion-exchange resin wafer type and ED/EDI operational parameters is obtained. Notably, it is found that the applied cell voltage has a predominant impact on both the separation efficiency and energy consumption. By employing multiobjective optimization techniques, experimental conditions that enable achieving 99% separation efficiency while keeping energy consumption below 1 kWh/kg are identified.