Machine Learning for Anion Exchange Membrane Used by Fuel Cells to Determine the Dominant Factor Controlling Anion Conductivity

Increasing global demand for clean and sustainable energy has increased attraction towards fuel cell for its zero emissive, sustainable and highly efficient properties. Current mainstream proton exchange membrane fuel cell (PEMFC) faces problems such as sluggish Pt-catalyzed oxygen reduction reaction (ORR) kinetics occurring at the cathode due to low pH operating environment of PEMFC, requiring more Pt to improve ORR reaction rate. In contrast, anion exchange membrane fuel cell (AEMFC) has better ORR reaction kinetics when compared to PEMFC due to the high pH operating environment of AEMFC. This potentially allows the reduction of Pt quantity used in AEMFC compared to PEMFC, thereby bringing down the cost of AEMFC. However, AEMFC carries problems such as low durability and low anion conductivity for the anion exchange membrane (AEM) used as electrolyte membrane in AEMFC, inhibiting it from commercialization. To date, AEM materials research and development (R&D) are carried out through experimental research, requiring huge amount of budget, time, and labor, thereby slowing down the progress of R&D. To solve such issues, high hopes are placed on utilizing machine learning (ML) to accelerate AEM R&D. However, readily available database regarding AEM materials, critical for implementing ML, are not available, and complex structure of AEMs made of functional polymers are difficult to represent in terms understandable by ML models. Hence, a database as well as a method to represent the complex structure of AEM polymers needs to be established to implement ML into the AEM materials research field.<br/>In this study, we sought to establish a database that contains both homopolymer and copolymer, represent it using descriptors understandable by ML models, and determine the dominant controlling factor of anion conductivity for AEM materials based on the chemical features information provided by the descriptors. AEM database was made by extracting papers listed in several review papers. Extracted polymer structures were converted to numerical form easily understandable by ML models. For machine learning, objective variable was set as anion conductivity, and explanatory variables were set as anion conductivity measuring temperature and polymer structure in numerical form. The models used in this study are chosen based on their ability to provide high prediction accuracy and difficulty in overfitting, which includes Category Boosting (Catboost), eXtreme Gradient Boosting (XGBoost), Gradient Boosting for regression (GBR) and so on. After training and validating the ML models, R² value was used to evaluate their prediction accuracy. Then, Shapley values for each model when predicting anion conductivity were calculated to determine the dominant controlling factor of AEM polymers.<br/>We successfully built a database containing 60 AEM papers worth of data, with 4,014 data points in it. Among the data points, there are 1,200 anion conductivity data points obtained from 180 types of AEM polymer structures. Machine learning models were trained in Python 3.9 using Intel Core i9-11900K. All trained models gave prediction accuracy exceeding 90%, with Catboost model achieving the highest prediction accuracy of 97.7%. To understand the prediction mechanism behind the machine learning models, Shapley values calculated from each ML models showed that anion conductivity measuring temperature carried the highest impact on predicted value. Others include polarizability of the polymer, meaning the polarizability of AEM polymer affects the anion conductivity of AEM heavily. Hence, during the design of AEM polymers, polarizability of the polymer should be increased to achieve higher anion conductivity. This shows that our model successfully determined the dominant control factor for anion conductivity of AEMs, carrying potential to serve as polymer design guidelines in the future.