InterPore2026

Comparison of CNN and GAN-Based Super-Resolution Methods for 3D Porous Microstructures

22 May 2026, 09:20

15m

Oral Presentation (MS15) Machine Learning in Porous Media MS15

Speaker

Rishabh Saxena (Helmut-Schmidt-Universität - Universität der Bundeswehr Hamburg)

Description

Across a wide range of energy and engineering applications, the performance of porous materials is strongly governed by their microstructure. In batteries, fuel cells, and hydrogen storage systems, microstructural features control key transport pathways and thus critically influence overall functionality. Accurate characterization therefore requires high-resolution (HR) three-dimensional (3D) microstructural data, since transport behavior depends heavily on fine-scale features. However, imaging methods such as focused ion beam–scanning electron microscopy (FIB-SEM) and X-ray computed tomography (CT) are costly and time-consuming, particularly at high spatial resolution.
To address these challenges, this work explores deep learning based super-resolution methods for generating HR 3D microstructures from low-resolution data. We study several super-resolution architectures, including CNN-based models (SRCNN, SRResNet, and U-Net) and a GAN-based approach (SRGAN). These 3D models take low-resolution inputs and reconstruct HR 3D microstructures. For comparison, we consider both geometric and transport properties: geometric fidelity is quantified using the Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR), while physical fidelity is evaluated by computing effective tortuosity and permeability via FEM solutions of the Laplace and Stokes equations, directly linking reconstruction quality to material functionality.
Deep learning based SR outperforms nearest-neighbor, bilinear, and bicubic interpolation; among the tested models, SRResNet best matches the ground truth in both structural and transport properties. SRGAN further shows that perceptual sharpness alone does not guarantee functional accuracy. Overall, evaluation on lithium-ion battery cathode materials indicates that deep learning models, particularly SRResNet, best preserve the key properties required for reliable HR microstructure reconstruction.