Reconstruction of Asphalt Nanostructures via Generative Adversarial Networks

Abstract

This study analyzed the influence of asphalt binder nanostructures and inherent variability on their mechanical responses. This was achieved by reconstructing asphalt binder nanostructures using generative adversarial networks (GANs). GANs are able to detect hidden patterns in a given data set automatically, and they generate distribution-free data because they are not bound to specific probability functions. The mechanical responses of GANs-generated nanostructures were analyzed using finite-element (FE) analysis. The asphalt nanostructures were captured using atomic force microscopy (AFM). To overcome the limited number of AFM nanostructures, image augmentation techniques and stochastic random fields (RF) modeling were used to generate virtual images of asphalts’ nanostructures, which were utilized as the expanded training data set for the GANs. This approach helped to enhance the GANs’ generalization abilities and avoid training problems such as mode collapse. Modeling parameters were optimized to expedite the training process and reduce the computational time. The results demonstrated that GANs are capable of generating probable arrangements of nanostructures, thus expanding the body of knowledge regarding probabilistic analysis of asphalt mechanical behavior. The ability of GANs to reconstruct nanostructures indicates that the complex and variable nature of asphalt nanostructures can be identified and replicated by GANs, offering the opportunity to generate many replicates of nanostructures without the need to conduct many laboratory tests. The results successfully related the properties and distribution of the asphalt nanostructures to the variation in the mechanical response. This study proposes a methodology that combines aspects of AFM testing, GANs modeling, and FE modeling for the analysis of the mechanical response of asphalt binder nanostructures. The outcomes of this work link nanoscale properties to bulk properties. Consequently, practitioners and researchers can use the developed methodology to study uncertainty in mechanical response of asphalt binders and design blends with tailored properties and enhanced performance. Researchers also can benefit from this methodology to incorporate sustainable modifiers that enhance fundamental nanoscale properties of asphalt binders. Moreover, this work enables computational evaluation of various material designs that would be impractical to conduct in a laboratory environment, thus conserving resources. Finally, the methodology can be generalized to study nanostructures of other engineering materials to optimize their performance.