Degree

Doctor of Philosophy (PhD)

Department

Chemical Engineering

Document Type

Dissertation

Abstract

Electrochemical systems are essential for sustainable water purification, energy generation, and material separations. However, their design and optimization are often hindered by nonlinear dynamics, limited datasets, and incomplete data. This dissertation addresses these challenges by developing hybrid modeling frameworks that integrate compositional modeling, molecular dynamics (MD), machine learning (ML), reinforcement learning (RL), transfer learning (TL), and machine learning potentials (MLP). These approaches bridge the gap between mechanistic and data-driven models, enabling accurate predictions, optimization, and scalable solutions for electrochemical systems.

The research begins with the development of an MD-ML framework to predict ion activity coefficients in ion-exchange membranes (IEMs), reducing reliance on extensive experimental data and eliminating the need for new fitting parameters typical of existing activity models. Next, a hybrid modeling approach for brackish water desalination via electrodialysis (ED) and electrodeionization (EDI) is introduced. This approach combines information from mathematical models with ML-based surrogate models to achieve high predictive accuracy and multi-objective optimization, enabling the design of conditions with high separation efficiency and low energy consumption. Additionally, an RL-based control framework is proposed, allowing for the autonomous optimization of operational parameters to improve ion removal efficiency.

Furthermore, a TL-based strategy is developed to address incomplete datasets in capacitive devices, combining data imputation with ML to enable accurate modeling and exploration of experimental conditions using multi-objective optimization. Finally, a physics-informed MLP framework is introduced to simulate ion transport in membrane-ion-water systems, using neural network potentials trained on ab initio molecular dynamics (AIMD) data for scalable and transferable modeling of structural and dynamic properties.

This dissertation presents novel methodologies that enhance data-driven modeling efforts in electrochemical separation processes. These methodologies demonstrate the effective incorporation of domain knowledge into data-driven models, facilitating a bidirectional flow of information between mechanistic and empirical models. This work advances the design, control, and optimization of electrochemical systems, paving the way for more efficient, scalable, and sustainable technologies.

Date

4-22-2025

Committee Chair

Jose Romagnoli

Download

Available for download on Saturday, April 22, 2028

Share

COinS