Graph Neural Network Modeling of Aggregate Contact Networks for Predicting Stress Distribution in Asphalt Mixtures

Abstract

Asphalt mixtures are complex particulate systems whose mechanical performance emerges from the arrangement, contact, and load transfer among coarse and fine aggregates, mastic, and voids. Predicting stress distribution within such materials is central to pavement design, durability assessment, and infrastructure sustainability, yet conventional continuum models often smooth over the heterogeneous contact networks that govern localized failure. This paper examines graph neural network modeling of aggregate contact networks as a system-level approach for predicting stress distribution in asphalt mixtures. It frames the mixture as a meso-scale infrastructure in which nodes represent aggregate particles or contact zones and edges encode physical adjacency, contact morphology, and load-transfer capacity. The discussion connects graph representation learning, discrete element modeling, image-based aggregate characterization, and materials informatics. It then analyzes architectural trade-offs involving inductive versus transductive learning, message passing depth, attention mechanisms, interpretability, and computational cost. The paper emphasizes robustness under noisy imaging, uncertain contact parameters, and distribution shift across mix designs and environmental conditions. It also considers deployment pipelines, data governance, fairness in infrastructure decision-making, and policy implications for transportation agencies. The central argument is that graph neural networks can provide a useful bridge between meso-scale contact topology and system-level stress prediction, but only when technical accuracy is coupled with validation, uncertainty quantification, sustainable computational practices, and transparent governance.