Physics-Guided Machine Learning Approach for Roughness Prediction in Jointed Plain Concrete Pavements

The International Roughness Index (IRI) is a crucial indicator of rigid pavement performance and distress, and artificial neural networks (ANNs) have made pavement management system (PMS) forecasting more cost-effective. However, purely data-driven ANN models are not constrained to follow known physical trends and can behave unpredictably outside their training range. This study developed a physics-informed neural network (PINN) for predicting IRI in jointed plain concrete pavements (JPCPs) by incorporating an empirical equation from the Mechanistic-Empirical Pavement Design Guide (MEPDG) into the training loss, intended to bias predictions toward physically consistent trends. A total of 2705 observations from 394 U.S. and 6 Canadian pavement sections were used to train the model and compare it against a conventional ANN, standard regression, and the standalone MEPDG equation. The PINN achieved a coefficient of determination (R2) of 0.695 and a root mean square error (RMSE) of 11.79 in./mi. Its performance was comparable to the ANN (R2 = 0.698, RMSE = 11.71 in./mi), and both outperformed the MEPDG equation and conventional regression. A closed-form equation was extracted from the trained ANN weights and biases. While the physics constraint did not improve predictive accuracy in this study, it offers a transparent, interpretable tool that agencies can apply directly to prioritize JPCP maintenance and rehabilitation without proprietary model files.