Micromechanics-Informed Neural Networks for Asphalt Mixture Stiffness and Ravelling Prediction

Abstract

Porous asphalt pavements are widely used on Dutch highways for their noise reduction and drainage benefits. However, they are susceptible to raveling. Raveling represents the progressive loss of surface aggregate that compromises both safety and structural integrity. Predicting when and where raveling will occur remains challenging because laboratory measurements of mixture stiffness and field observations of surface deterioration have traditionally been treated as separate problems. This thesis develops micromechanics-informed machine learning frameworks that connect laboratory mixture characterization to field performance prediction. The research proceeds in two main stages. In the first stage, Micromechanics-Informed Neural Networks (MINNs) are developed for predicting asphalt mixture stiffness from composition and constituent properties. The networks embed established micromechanical theories as physics constraints within the learning objective. Mori-Tanaka and Differential Scheme homogenization serve as the primary theoretical foundations. A systematic comparison against Modified Hirsch baselines using 371 laboratory samples demonstrates that the Eshelby-based MINNs achieve superior predictive accuracy (𝑅2 = 0.859 for Mori–Tanaka) while maintaining strong adherence to physical principles. The investigation also reveals that adding more physics constraints does not automatically improve performance; a two-term loss formulation outperforms a threeterm formulation that includes explicit boundary conditions, highlighting the importance of optimisation design in physics-informed learning. In the second stage, the framework extends from laboratory stiffness to field raveling prediction using monitoring data from Dutch highways. This extension requires fundamental adaptations: physics parameters must be discovered from observational data rather than computed from theory, and hierarchical encoding architectures are needed to manage the uncertainty inherent in field measurements. The temporal analysis reveals a key finding termed the Goldilocks Zone. Specifically, physics-informed methods provide maximum benefit at intermediate data volumes ranging from 4 to 6 years of monitoring. At such intermediate volumes, substantial improvements over baseline models are observed. However, the benefits are moderate to incremental as datasets grow larger. Ultimately, the finding provides practical guidance for when physics integration is most valuable and when simpler approaches suffice. The main contributions of the thesis include three key elements. First, validated MINN architectures enable frequency-dependent stiffness prediction. Second, data-volume thresholds governing physics-informed learning effectiveness are identified. Third, practical deployment guidelines for pavement management agencies are provided. The work advances both the methodology of physics-informed machine learning and its application to infrastructure management. Furthermore, the research establishes a foundation for predictive maintenance systems. Such systems combine the pattern-recognition capabilities of data-driven learning with the interpretability of physics-based domain knowledge.