Prediction of International Roughness Index (IRI) Using Physics-Informed Neural Networks (PINN)

Abstract: The International Roughness Index (IRI) is one of the most vital indicators used in The Pavement Management System (PMS). Prediction of IRI of flexible pavements pivots the way of sustainable and cost-effective strategies for pavement management. With the help of Artificial Neural Networks (ANN), the IRI can be forecasted before the deterioration of pavements. Though neural networks mimic the brain's cognitive process for learning its function complexity sets it apart from other empirical methods. The study aims to bridge the neural networks with physics-based prediction methods. This novel approach of utilizing the physics-based equation for neural network training offers a more reliable prediction of IRI. The Study demonstrated a framework for the Physics Informed Neural Networks (PINN) model. This ANN model integrates an equation from the Mechanistic-Empirical Pavement Design Guide (MEPDG) into its training regimen. The input parameters followed the physics-based equation. A total of 390 data points of 74 flexible pavement sections from the United States and 8 pavement sections from Canada were used to train the model. The data points were extracted from the Long-Term Pavement Performance (LTPP) database. The neural networks model has demonstrated promising outcomes, attaining an accuracy exceeding 80% in terms of IRI prediction. Furthermore, the Sensitivity Analysis of the model input parameters was evaluated and showed a positive correlation with the model output. This model framework can be adopted by transportation management agencies for predicting IRI to efficiently allocate resources and prioritize maintenance activities.