Leveraging physics with deep learning: physics-informed neural networks (PINN) for IRI prediction in flexible pavements

Abstract

This study presents a physics-informed neural network (PINN) framework for predicting the international roughness index (IRI) of flexible pavements, utilizing the long-term pavement performance database. A total of 390 observations from 74 pavement sections across the United States were utilized. The architecture of the PINN combines a mean square error (MSE) loss function with a custom physics-informed MSE (PMSE) loss function. It integrates mechanistic IRI equations sourced from the mechanistic-empirical pavement design guide (MEPDG) into the training process. Comparative analysis was performed against standard regression models and the MEPDG IRI prediction equation. Results indicate that the PINN model significantly improves prediction accuracy, achieving a coefficient of determination ( R 2 ) of 0.743 and a root mean squared error of 26.57, outperforming traditional methods. Furthermore, a closed-form equation from the model was derived. By utilizing this new simplified equation, transportation agencies can allocate resources and prioritize maintenance activities efficiently through the IRI forecast PINN Model for asphalt pavements.

Topics & Concepts