Adaptive sampling and artificial neural networks for surrogate modelling in high-dimensional mechanistic-empirical rigid pavement design and analysis

AASHTOWare Pavement ME Design (Pavement ME) is a software tool for implementing Mechanistic-Empirical (M-E) design for rigid pavements. However, the difficulty in quantifying many design inputs and the time-consuming trial-and-error design process make it computationally expensive and less efficient for pavement design practice. To overcome this, machine learning (ML)-based surrogate models were developed to accelerate the M-E design process. Nevertheless, their development typically requires a large volume of Pavement ME simulation data, which can quickly become intractable due to the limitation of time and computational resources, especially for high-dimensional modelling. This paper bridges the gaps between high-dimensional surrogate modelling and their efficient development using a new sampling technique, named scalable adaptive sampling (SAS). Several artificial neural network (ANN) models were developed for different climate regions with typical pavement structures and a wide range of many design variables, including material properties, portland cement concrete (PCC) slab geometry, traffic configurations and loads, as well as pavement age. The developed ANN models were able to replicate Pavement ME predictions of fatigue damage and long-term transverse cracking for rigid pavements with high accuracy, while significantly improving computational efficiency. The developed ANN-based surrogate models facilitate the high-dimensional jointed plain concrete pavement (JPCP) design consideration while still achieving speed and accuracy.