Pavement Cracking Prediction Models Based on Deep Learning Physics-Informed Neural Network

Abstract

This study investigates the application of physics-informed neural networks (PINNs) to predict pavement deterioration, specifically, cracking in the wheel path in asphalt concrete pavements, leveraging the Federal Highway Administration long-term pavement performance (LTPP) database. PINNs attempt to address the black-box nature of conventional deep learning models by integrating physical rules directly into the model architecture for prediction. The study employs comprehensive pavement data from the LTPP program, focusing on factors such as construction quality, traffic loads, environmental factors, and maintenance-related changes to the structure. Key findings are summarized, including the improved predictive accuracy and generalization capabilities of PINN models. Enforcing an expected positive correlation of pavement age versus cracking also increased prediction accuracy in comparison to conventional artificial neural network (ANN) results. Specifically, the fifth model that applied a positive correlation with ESAL, and a negative correlation with the number of layers, achieved the highest R² and the lowest mean absolute error on the testing datasets. The results underscore the importance of incorporating engineering knowledge into pavement performance models, particularly, features such as ESAL, pavement age, layer structure, and precipitation. The incorporation of these domain-specific features in PINN models resulted in more accurate and robust predictions of wheel path cracking for asphalt concrete pavement.