Combined Materials Map and Supervised Machine Learning Towards Efficient and Accelerated Anion Exchange Membrane Polymer Exploration

Decarbonization requires next generation energy generating devices and fuel sources to replace fossil fuel-powered infrastructures. Fuel cells and water electrolyzer are one such candidate to realize decarbonized society, and anion exchange membrane (AEM) is seen as the hopeful to aid their widespread adoption because of its potential to suppress usage of expensive catalyst. However, commercialization of AEM requires improvement in anion conductivity and durability. Many researches have been done to date, with progress being made for the past 20 years. Regardless, AEM that fulfill both high anion conductivity and high durability is yet to be found, and exploration for such material needs to be accelerated for early decarbonization. This is where materials informatics comes into play, with the advent of prevailing machine learning (ML) empowering a transformative opportunity to accelerate AEM material development. ML algorithms can be broadly classified into supervised and unsupervised learning. Supervised ML models seeks for a correlation between target variables (e.g., anion conductivity) and explanatory variables (e.g., polymer structure information) based on labeled data. In contrast, unsupervised ML models explore the relationships among explanatory variables without relying on target variables, enabling the identification of latent patterns and features. Building upon our previous work on supervised explainable ML model [1] and unsupervised ML-based material mapping [2], we present an integrated approach that utilizes unsupervised ML-derived 2-dimensional (2D) map as material design guidelines for identifying promising new AEMs. Through this method, novel AEMs can be designed easily through reflecting knowledge obtained from the guideline. These AEMs are then subjected to supervised ML-based conductivity prediction.<br/>Since an opensource AEM database is currently unavailable, we constructed our own database from 78 research papers. The Mordred library [3] was employed for descriptor transformation of AEM polymer structures. The 2D maps were generated by combining principal component analysis (PCA) and uniform manifold approximation and projection (UMAP). CatBoost was employed for supervised ML. Two types of descriptors were used to train CatBoost: AEM polymers represented using Mordred descriptors and dimensionally reduced descriptors using PCA (PCA descriptor, 32 dimensions). The predictive accuracy of CatBoost was evaluated using the coefficient of determination (R²). Both descriptor sets yielded high cross-validation accuracy, marking R² values of 0.961 and 0.957, respectively. To assess the practicality of our integrated approach, we employed poly(2,6-dimethyl-1,4-phenylene oxide) (PPO)-based AEMs [4] with known ionic conductivity as pseudo new AEM polymer. Assuming the case that this new AEM has not been reported yet, we could have thought of designing it through utilizing piperidinium, a side chain that provided high average anion conductivity to poly(arylene alkylene), with PPO as main chain structure due to its ease of synthesis, all thought of through utilizing the map as guideline. Next, the anion conductivity for the AEM measured at different temperatures was accurately predicted by CatBoost for the test AEM structure, even when represented using either sets of descriptors (Mordred descriptor and PCA descriptor), with R² values of 0.906 and 0.888, respectively. Our work demonstrates the remarkable potential of integrating data-driven material mapping and ML for accelerating AEM development. This synergistic approach has the potential to revolutionize AEM research, enabling the rapid identification and design of high-performance AEMs for next-generation AEM fuel cells.<br/>[1] Y.K. Phua et al., <i>Sci. Technol. Adv. Mater.</i>, 2023, 24, 1, 2261833. [2] Y.K. Phua et al., <i>ChemElectroChem</i>, 2024, in press. [3] H. Moriwaki et al., <i>J. Cheminformatics</i>, 2018, 10, 4. [4] J. Chen et al., <i>Ind. Eng. Chem. Res</i>., 2022, 61, 4, 1715-1724.