Machine-Learning-Driven Stochastic Modeling Method for 3D Asphalt Mixture Reconstruction from 2D Images.

Three-dimensional reconstruction programs are essential tools for understanding the behavior of asphalt mixtures. On the basis of accurate 3D models, it is convenient to identify the complex relationship between spatial structures and physical properties. In this work, we explore a low-cost and data-efficient way to create a collection of 3D asphalt mixture models. The core idea is to introduce a foundational segmentation program and stochastic modeling into the asphalt mixture reconstruction framework. First, our approach captures a 2D image to present spatial structures of the investigated sample. The integration of a smartphone camera and an image quilting method has been designed to understand fine-grained details and facilitate full coverage. Aiming at realizing high-quality segmentation, we propose the Segment Anything Model (SAM)-driven method to distinguish aggregate grains and asphalt binder. Second, Multiple-Point Statistics (MPS) is activated to build 3D models from 2D training images. To speed up the reconstruction step, we apply Nearest Neighbor Simulation (NNSIM) to improve pattern searching efficiency. Aiming at calculating 3D conditional probabilities, the probability aggregation framework is introduced into the asphalt mixture investigation. Third, our program focuses on the modeling evaluation procedure. Determination of a two-point correlation function, analysis of distance and a grain size distribution assessment are separately performed to check the reconstruction quality. The evaluation results indicate that our program not only preserves spatial patterns but also expresses uncertainty during the material production step.