Roughness Prediction of Jointed Plain Concrete Pavement Using Physics Informed Neural Networks

Abstract

The International Roughness Index is used to measure the road roughness in pavements, as pavement roughness deteriorates over time. Despite many attempts by researchers to predict roughness in concrete pavements, there are limitations, such as small sample size, modeling approach, or lack of robustness in the model. This study presents a novel machine-learning approach incorporating domain knowledge to predict roughness, using a dataset obtained from the Long-Term Pavement Performance database. Physics informed neural networks (PINNs) are popular physics-driven machine-learning approaches that have been receiving widespread attention in the field of civil engineering. PINNs work similarly to neural networks but are augmented with the incorporation of physics-based constraints and governing equations, enabling them to assimilate domain knowledge and leverage physical principles while making predictions or solving problems. In this study, the popular Mechanistic-Empirical Pavement Design Guide roughness prediction model is used along with the optimized neural networks to calculate the physics-based loss function. The Optuna framework is used to tune the hyperparameters within the neural network architecture. The final configuration, optimized and trained in the model, has three hidden layers with, respectively, 27, 67, and 80 neurons. The tuned model has performed well for the testing dataset, with a mean absolute error of 0.134 and a coefficient of determination of 0.90. A sensitivity analysis was also conducted and is presented to understand the influence of the variation of each variable.