ABSTRACT Pavement surface smoothness (or roughness) is crucial for traffic safety, driving comfort, and fuel efficiency. As a widely applied roughness indicator, an accurate forecasting of the International Roughness Index (IRI) and its deterioration is essential for the design, maintenance, and management of asphalt pavements. Previous studies have used field measurement data or AASHTOWare Pavement ME Design simulations for the development of machine learning (ML) models to streamline the IRI modelling. However, these models frequently lack the accuracy and robustness of the measurement data or high-fidelity computational simulations they are intended to surrogate. To address this issue, we employed a new adaptive sampling technique to generate an informative yet efficient pavement damage database from Pavement ME simulations. Utilising Artificial Neural Networks (ANNs), we engineered two types of surrogate ML models: (a) Model I, an ANN-based IRI predictive model, and (b) Model II, a hybrid model combining ANN-based predictions of rutting, fatigue damage, and thermal cracking with closed-form relationships between these indicators and IRI. Our findings show that Model II outperforms Model I in IRI modelling accuracy both globally and locally. Moreover, Model II matches IRI simulations of Pavement ME while providing enhanced efficiency and adaptability to a broader spectrum of design considerations.
Read full abstract