Data-driven automated modeling technology for precision mesostructure reconstruction of asphalt concrete

The advancement of road research demands higher efficiency and refinement in the modeling of asphalt concrete (AC) mesostructures. Computed tomography (CT)-based modeling can more accurately reflect the real mesostructure and generally involves two-dimensional (2D) image segmentation and three-dimensional (3D) reconstruction. However, existing 2D image segmentation methods generally suffer from weak generalizability and low processing efficiency. To address this issue, a full-angle CT data expansion method based on real mesostructures was developed to increase the quantity and diversity of the training data. Subsequently, an AEUnet-DIP proxy model based on an Attention-Enhanced U-Net (AEU-Net) and digital image processing (DIP) was developed to achieve efficient and accurate segmentation of AC CT images. In addition, during 3D reconstruction, independently processing individual 2D CT images often neglects the geometric correlation between adjacent CT images, which may lead to randomly distributed structural defects in the reconstructed mesostructure. To address this problem, the CT Image Full-angle Correction Modeling Program was developed based on the geometric correlation between adjacent CT images and coordinate-space transformation matrices. Finally, the proposed methods were analyzed and evaluated. The results showed that the AEUnet-DIP proxy model can efficiently and accurately segment AC CT images and demonstrates strong generalizability across CT images with different sources, geometries, and image characteristics. Meanwhile, the CT Image Full-angle Correction Modeling Program effectively reduces spatially distributed structural defects while preserving mesostructural details, achieving a favorable balance between defect elimination and geometric fidelity. The resulting mesostructural models are well suited for subsequent thermal and mechanical numerical simulations.