Reconstructing 3D Digital Twins of Aerogels using Generative Adversarial Networks

Abstract

Aerogels are highly porous materials created from gels by replacing the liquid with air using a drying process and have porosity values of close to 90 to 99 %. They have excellent applications in aerospace, automative and other industries due to their high thermal insulation properties.The property of these class of materials are governed predominantly by their nanoporous micro structures. Due to this microporous structure, it is a significantly complex problem to accurately estimate their structure-property relationships, which is critical in developing aerogels for targeted applications. Thus, determining the physical properties of aerogels is an important step in their utilization for a wide range of industrial applications. To this end, a digital twin of such aerogels would aid us in analysing and understanding the physical properties needed for industrial applications in a safe and energy efficient fashion. This research aimed to develop computational approaches to reconstruct the 3D microstructure of a silica-based aerogel using Conditional Generative Adversarial Networks and Convolutional Neural Networks (CNNs), where the CNNs would predict the structural properties of the 3D aerogels through synthetic Scanning electron microscopy (SEM) images created from their 3D structure. To this end CNNs that predict the four structural properties of radius,relative density, pore size distribution (mean and standard deviation) were developed which predicted these properties with minimal errors. A 3D Conditional Generative Adversarial Network was developed which is able to capture spherical information from the dataset without conditioning and reconstructs the underlying structure of the 3D dataset without the spherical information when conditioned with the relative density property of the aerogel, which is predicted by the CNN.