Machine Learning for Understanding Microstructures and Morphologies from Materials for Ionomer-Based Water Electrolysis

Integrating experimental data and simulations is critical for advancing data-driven algorithms based on machine learning to create new applications that describe material behaviors. Our research focuses on two main objectives: investigating microstructures from experimental data and enhancing the connection to simulation models for ionomer-based water electrolysis. By examining the intricate morphologies of materials, this work aims to bridge the gap between experimental observations and computational predictions, moving towards a digital twin. Current research at Berkeley Lab includes establishing baselines for interfacial studies in ionomers, examining interfaces in both acidic (e.g., proton-exchange, Nafion) and alkaline environments. Evaluating catalysts is a key consideration, focusing on iridium for the cathode while exploring less costly alternatives. Material formulation involves inspecting catalyst ink, comprising polymer ionomer, nanoparticles, and other components. Imaging and analysis opportunities include monitoring structural changes at material interfaces due to contact with adjacent materials and assessing degradation issues like ionomer decomposition influenced by potential. Our understanding of the length and timescales of experimental data, alongside detailed knowledge of surface structure, chemistry, and morphology, depends upon using state-of-the-art computer vision (e.g. CNN, vision transformers and hybrids) and natural language processing (e.g. large language models for topic modeling) techniques. Modeling and predicting material properties from microstructures to macroscopic behaviors is essential for designing and/or verifying simulation algorithms that interpret complex data and adapt to various scales, from atomic-level interactions to visible surface patterns. Bringing simulation data to empirical observations in a unity of theory and experiments to interrogate and interpret experiments requires advanced imaging techniques to optimize promising designs. This integration is also vital when using electron and X-ray spectroscopic and imaging data for material characterization. For example, understanding how electrochemical reactions at the electrode-electrolyte interface modify surfaces can significantly influence models. Current investigations about porous transport layer (PTL)-membrane interactions and their deformation monitored in situ using microCT highlight the importance of detailed morphological studies using multimodal information for resolving these soft-hard interfaces for improved function. By focusing on the length and timescales of materials and respective structures, this research seeks to improve the accuracy of models and material predictions. This approach enriches characterization algorithms, enabling them to interpret complex data more effectively and adapt to various scales. This synergy will refine existing computational representations and help pioneer new models capable of predicting the behavior of complex materials. Presenting this work at the Symposium MT02: Machine Learning in Action—Automated and Autonomous Experiments aims to contribute to the dialogue on the potential of machine learning in material science and engineering. This work was supported by the U.S. Department of Energy, Office of Science Energy Earthshot Initiative, as part of the Center for Ionomer-based Water Electrolysis under contract/grant #DE-AC02-05CH11231.