A Diffusion-Based Digital Twin for Predicting Road Deterioration under Climate and Traffic Stress

Accurately tracking the spatiotemporal evolution of road distress is essential for formulating informed repair and maintenance strategies. However, existing road performance prediction approaches typically rely on weighted composite indices such as the Pavement Condition Index (PCI) rather than fine-grained, distress-level observations. This coarse representation limits the ability to detect early-stage deterioration and constrains the development of precise, predictive maintenance models. To address these limitations, this study proposes a physics-embedded diffusion (PED) model that provides a physically consistent spatiotemporal framework for jointly modeling distress evolution, maintenance interventions, and environment-induced disturbances. Validation using real-world datasets from China demonstrates strong agreement between PED predictions and observations, achieving a mean PCI mean absolute percentage error (MAPE) of 9.262% and a mean weighted percentage error of area (area MWPE) of 4.484%. Furthermore, evaluation on the US Long-Term Pavement Performance dataset confirms the model's strong cross-regional generalization capability, achieving a mean PCI MAPE of 28.054% and an area MWPE of 8.957%, underscoring the robustness and transferability of the proposed framework. Experimental results further indicate that the optimal historical environmental window for short-term distress prediction is approximately 50 d. As the volume of training data increases, the model continues to maintain stable convergence performance, demonstrating the scalability and robustness of the proposed physics-guided deterioration modeling approach. This scalable tool is invaluable for simulating realistic road deterioration processes, providing a reference for evaluating subsequent maintenance decision-making strategies.