Physics-Informed Machine Learning Models for Predicting Pavement Deterioration in Smart Highway Networks: A State-of the-Art Review

Abstract

ABSTRACT Accurate prediction of flexible and rigid pavement deterioration is vital for optimizing Pavement Management Systems (PMS) and ensuring sustainable smart highway infrastructure. While traditional Mechanistic-Empirical (M-E) pavement design frameworks suffer from restrictive physical simplifications and calibration overhead, purely data-driven Machine Learning (ML) techniques—such as Deep Neural Networks (DNNs), Random Forests (RF), and Extreme Gradient Boosting (XGBoost)—frequently produce physically inconsistent forecasts, extrapolate poorly beyond training domains, and require massive sensor datasets. Physics-Informed Machine Learning (PIML), particularly Physics-Informed Neural Networks (PINNs), has emerged as a transformative paradigm that seamlessly embeds domain-specific physical principles—including viscoelastic constitutive laws, damage accumulation theories (e.g., Miner's cumulative damage hypothesis), dynamic fatigue equations, and climate-induced thermal transport partial differential equations (PDEs)—directly into loss functions during neural network training. This paper presents a comprehensive, journal-level critical review of PIML architectures applied to pavement deterioration modeling within smart highway networks integrated with Internet of Things (IoT) and Weigh-In Motion (WIM) sensors. We systematically examine governing mechanics based formulations, data-driven surrogate models, hybrid integration topologies (e.g., residual learning, physics-guided architecture design, exact hard-constraint enforcing), and multi-objective loss optimization algorithms. Comparative benchmarks reveal that PIML models achieve up to 42% reduction in Root Mean Square Error (RMSE) under sparse data regimes (<15% sensor coverage) and maintain strict physical plausibility during 20 year service life extrapolation. Finally, key technological challenges— including non-convex multi-objective loss landscapes, sensor noise, real-time edge computing constraints, and model interpretability—are critically evaluated, followed by strategic directions for future smart transportation infrastructure research.

KEYWORDS: Physics-Informed Machine Learning (PIML); Pavement Deterioration; Physics-Informed Neural Networks (PINNs); Smart Highway Networks; Pavement Management Systems (PMS); Constitutive Degradation Laws; Structural Health Monitoring (SHM).