will defend a doctoral thesis entitled,

AI-Accelerated Design of Anion Exchange Membranes

On

Tuesday, November 17 at 9:00 a.m.
MRDC Room 3515

and

 Virtually via MS Teams

Teams Link

Committee
            Prof. Rampi Ramprasad – School of Materials Science and Engineering (advisor)
            Prof. Ryan P. Lively Doe– School of Chemical and Biomolecular Engineering

      Prof. Guoxiang (Emma) Hu– School of Materials Science and Engineering

      Prof. Scott Danielsen– School of Materials Science and Engineering

      Prof. Chao Zhang– School of Computational Science and Engineering

Abstract
Anion exchange membrane (AEM) fuel cells could replace costly platinum catalysts and fluorinated polymers in proton exchange membrane fuel cells, but AEMs remain limited by low ionic conductivity and rapid degradation in alkaline media. This dissertation develops a data-driven framework to accelerate AEM design. First, multi-task machine learning models trained on a literature-curated dataset predict hydroxide conductivity, water uptake, and swelling ratio from polymer structure and operating conditions; screening approximately 11 million candidate copolymers identified 478 fluorine-free formulations predicted to meet targets for all three properties. Second, a physics-enforced neural network (PENN) predicts four physically meaningful degradation parameters for each membrane, revealing a universal degradation curve across chemistries and conditions and forecasting conductivity over thousands of hours from approximately 200 hours of early-time data. Third, because both models were limited by the size and diversity of available data, SALSA (Semi-Autonomous Literature Summarization Assistant) was developed: an open-source, human-in-the-loop platform that extracts structured data from the text, tables, and figures of scientific documents while preserving the evidence behind every value. Compared with expert manual curation, SALSA halved curation time at a total model cost of $10.82 for 25 publications, and it was used to expand the AEM dataset to 9,535 publicly released data points from 121 publications. Retraining the PENN on this expanded dataset confirmed the universal degradation curve for more than twice as many material groups and improved predictions for unseen publications. Preliminary work shows that generative AI can rapidly build agents for data curation and molecular simulation tools, but that their outputs still require careful validation. Together, these contributions establish a closed loop in which curated data improve the models and the models identify the data needed next, providing a blueprint for data-driven design of sustainable polymer electrolytes and other complex materials.