3D Microstructure Reconstruction of Aerogels via Conditional GANs

Abstract

Aerogels are low-density and highly porous materials (90–99% porosity) with exceptional thermal and mechanical properties, governed by their intricate nanoporous microstructure. Understanding their structure-property relationships is essential for optimizing their performance across industrial applications. A sig- nificant challenge appears in precisely identifying the complete pore space and thus mapping their microstructural morphology of aerogels. This work presents a deep learning-driven digital twin framework for aerogels, leveraging Conditional Generative Adversarial Networks (cGANs) and Convolutional Neural Networks (CNNs) for 3D microstructure reconstruction and predictive modeling. Our ap- proach reconstructs 3D aerogel microstructures from synthetic 2D scanning elec- tron microscopy (SEM) images that mimic real samples by incorporating depth effects. A CNN predicts key microstructural parameters, including pore radius, relative density, and pore size distribution, with minimal error. A 3D cGAN then generates aerogel microstructures by capturing global spatial features and condi- tioning on the extracted parameters. We demonstrate that conditioning improves the fidelity of reconstruction by en- forcing physically meaningful constraints. This method provides a scalable, data- driven approach for microstructure modeling, enabling efficient structure-property predictions, and guiding aerogel design for targeted applications.