Abstract
Asphalt concrete (AC) is a heterogeneous aggregate–mortar material whose effective mechanical properties are strongly governed by mesostructural geometry and constituent material properties. Establishing quantitative relationships between real mesostructures and the elastic modulus is important for reducing experimental cost and improving material design efficiency. This study proposes an image-driven forward prediction and inverse analysis framework for elastic modulus evaluation of AC based on real mesostructural images. A mesostructural image sample generation algorithm is developed to expand realistic datasets derived from computed tomography (CT) images, and an automated image-based finite element (FE) modeling framework is established to generate elastic modulus labels. By integrating real mesostructural geometry, constituent material properties, and FE analysis, an offline mesostructural image–elastic modulus database is constructed. Based on this database, a convolutional neural network (CNN)-based prediction method is developed to efficiently estimate the elastic modulus of independent AC mesostructures, while a gradient-based inverse analysis method is introduced to identify key mesoscopic geometric contributors. The results show that the proposed framework enables efficient mapping from real mesostructural images to the elastic modulus and provides interpretable information on the effects of global and local mesostructural variations. By reducing reliance on repeated specimen preparation and mechanical testing, the proposed method can decrease material consumption, testing costs, and experimental waste, thereby supporting cleaner and more resource-efficient AC material design.