Prediction of Gas Diffusion Coefficients Using Manifold Learning and X-ray Ptychography Data

A key challenge in advanced materials science is establishing quantitative structure–property relationships for complex hierarchical architectures spanning micrometer to nanometer scales. Predicting gas diffusion coefficients in the catalyst layer (CL) of the proton exchange membrane fuel cell (PEMFC) catalyst-coated membrane (CCM) exemplifies this challenge. The CL’s intricate porous network governs PEMFC performance, durability, and cost. Although recent advances in hard X-ray ptychographic nano-computed tomography (nano-CT) enable nondestructive visualization of such complex structures with nanometer resolution, constructing a quantitative structure–property relationship remains a major bottleneck.

Here, we present a manifold-learning-based framework that directly links the three-dimensional porous structure to diffusion properties of the CL. This unsupervised approach implicitly extracts intrinsic structural features without manual feature engineering, enabling robust property prediction from structural data alone. The model predicts gas diffusion coefficients with relative errors below 10%, even for low-resolution data where direct diffusion simulations fail. Validation using high-resolution X-ray ptychographic nano-CT data acquired at the NanoTerasu facility demonstrates the framework’s reliability. This work establishes a new paradigm for data-driven materials design targeting complex hierarchical systems.