Kyushu University establishes new approach for efficient materials development

A research group, including Professor Koichiro Kato and Professor Tsuyohiko Fujigaya from the Faculty of Engineering, along with third-year Doctoral Student Phua Yin Kan from the Graduate School of Engineering at Kyushu University, has developed a "human-in-the-loop" framework. Designed for anion exchange membrane (AEM) materials, which serve as core components in fuel cells and water electrolyzers, this framework combines explainable AI, ChatGPT, and expert knowledge.

With this method, the group successfully established a quantitative link between structure and material properties at the molecular descriptor level, a process that previously relied heavily on rules of thumb. They extracted quantitative design guidelines that experimental researchers can readily use, such as the effectiveness of a biphenyl backbone and the importance of a side-chain length spanning eight bond distances.

This approach allows researchers not only to see what the AI predicts but also to understand why it made that prediction, enabling them to apply these insights directly to materials design. This breakthrough is expected to reduce trial and error in materials development and streamline the screening of candidate materials. The findings were published in the Journal of Materials Chemistry A.

Anion exchange membranes (AEMs) are core components of fuel cells and water electrolyzers, making them vital materials for achieving a hydrogen-based society. In the molecular design of AEM polymers, developers must balance high ion conductivity with long-term alkaline stability. However, these are competing requirements, and design choices have continued to rely largely on empirical rules of thumb.

While machine learning, particularly Artificial Neural Networks (ANNs), can predict the properties of polymer materials with high accuracy, the bases for their predictions are notoriously difficult to understand. This "black box" nature has made it challenging for experimental researchers to apply prediction results to actual molecular design. Furthermore, explainable AI (XAI) methods require immense computational costs when applied to ANN models dealing with high-dimensional molecular descriptors, making them difficult to use for real-world material systems.

Previously, the research group built an original database systematically cataloging the chemical structures and properties of 346 types of AEM polymers. They used unsupervised machine learning to create materials maps, visualizing the relationships between structure and performance from a bird's-eye view. However, it remained difficult to quantitatively extract relationships between individual structural features and physical properties from material maps and to present them as practical design guidelines for synthesis strategies of experimental researchers.

In this study, the group used their unique database to construct a framework that extracts quantitative molecular design guidelines that experimental researchers can use from a black-box ANN model.

First, using a unique two-step dimensionality reduction strategy based on statistical methods and explainable AI, they compressed a molecular descriptor space of several thousand dimensions down to just 67 dimensions. This reduction allowed them to apply an explainable AI analysis method called SHAP to the high-dimensional ANN model, which was previously hindered by computational cost barriers, while simultaneously improving the prediction accuracy of the ANN itself.

Next, the group utilized ChatGPT to help interpret the critical descriptors identified by the SHAP analysis. By providing ChatGPT with the source code and official definitions of the descriptors, they used the AI to help translate mathematically abstract descriptors into chemically intuitive language.

Because ChatGPT's outputs exhibited certain biases and limitations in interpretation, the group incorporated a verification and correction process by human experts to ensure the reliability of the interpretations.

Through this workflow, the group successfully extracted quantitative molecular design guidelines. These included the effectiveness of the biphenyl backbone, the importance of maintaining a side-chain length of eight bonds from the main chain to the cation site, and the significance of introducing a heteroatom every four carbons. These guidelines align independently with the latest experimental findings, providing a quantitative, descriptor-level foundation for design insights that had previously been discussed only as empirical rules of thumb.

Furthermore, of four conceptual AEM polymers proposed based on the extracted design guidelines, two of them were predicted to exhibit an anion conductivity of 0.1 S/cm or higher at 80℃.

Given that only 14.4% of the polymers in the training database met this condition, the result suggests that this framework is highly effective for efficiently narrowing down promising candidates.

Because this framework can be applied generally to any material system that uses molecular descriptors, it can extract design guidelines that experimental researchers can understand and utilize from black-box AI predictions across a wide variety of functional polymer materials.

It is expected to reduce trial and error in materials development, leading to a reduction in the time and cost required for synthesis and evaluation.

Moving forward, the research group will collaborate with Associate Professor Manabu Tanaka of Tokyo Metropolitan University, who is a co-author of the English article, to proceed with the experimental synthesis and evaluation of the conceptual polymer structures proposed by the framework. This will experimentally verify both the prediction accuracy and the validity of the design guidelines.

The database and source code have been made publicly available on GitHub, raising expectations for other research groups to deploy and verify the framework on alternative materials.

Journal Information
Publication: Journal of Materials Chemistry A
Title: Orchestrating explainable AI, ChatGPT, and human expertise: a framework for extracting polymer design guidelines
DOI: 10.1039/D5TA06120B