Machine Learning for Prediction and Synthesis of Anion Exchange Membranes

Abstract

Anion exchange membrane fuel cells (AEMFCs) and water electrolyzers (AEMWEs) play a crucial role in the utilization and production of hydrogen energy, offering significant potential for widespread application due to their high energy conversion efficiency and cost-effectiveness. Anion exchange membranes (AEMs) serve the dual purpose of gas isolation and the conduction of OH^– ions. However, the poor chemical stability, low ionic conductivity, and inadequate dimensional stability of AEMs hinder the development of AEM-based energy devices. AEMs exhibit a more intricate chemical structure than general polymers, primarily due to their complex composition and unique attributes. This complexity is attributed to varying chain lengths, degrees of branching, and copolymerization compositions. Furthermore, diverse ion types, ion distribution, ion exchange capacity, hydrophilic clusters, electrostatic interactions, and microphase morphology further complicate these characteristics. In the past decade, more than 5,000 references have been dedicated to obtaining high-performance AEMs. Despite the large amount of work conducted during this period, the performance of AEMs still falls short of meeting the actual needs. The trial-and-error method used in designing membrane structures has proven inefficient and costly. Machine learning, a data-driven computational method, leverages existing data and algorithms to predict yet-to-be-discovered properties of materials. Recently, our group and some researchers have utilized machine learning to expedite the process of material discovery and achieve accurate synthesis of high-performance AEMs.

In this Account, we summarize the state-of-the-art for the AEMs, encompassing the structure design of cations and polymer backbones, strategies to improve the ion conductivity, and challenges arising from the necessity to achieve a delicate equilibrium among high conductivity, alkaline stability, and dimensional stability. Furthermore, we conduct a comprehensive review of recent breakthroughs in machine learning, specifically analyzing their implications within the context of AEMs. We examine the two primary branches of machine learning, supervised and unsupervised learning, and summarize various machine learning models, discussing the applicability and limitations of different algorithms. It is particularly worth noting that machine learning has the capability to predict the various properties of AEMs, such as conductivity and alkaline stability, and it can even design the structure of AEMs in accordance with the specific performance requirements of energy devices. By effectively screening high-performance membrane structures from millions of unknown candidates, machine learning significantly reduces the development time and cost associated with AEMs. Consequently, this technological advancement accelerates the rapid progress of AEM-based energy devices. Finally, we highlight the current challenge and future potential for machine learning to enable the development of superior AEMs.