Artificial Neural Networks for Mechanistic-Empirical Concrete Pavement Design: An Efficient Development Approach

Abstract

The AASHTO Mechanistic-Empirical (M-E) pavement design method offers cost-effective and sustainable design solutions for concrete pavements. However, its implementation is hindered by the complexity and time-consuming nature of the iterative computational process required to obtain optimal designs using the companion Pavement ME software. Furthermore, high license fees limit its wider adoption. Previous attempts to employ machine learning (ML) models for M-E design, instead, faced challenges in balancing model accuracy and capability with the increasing number of required training samples when including more design variables. Consequently, existing ML-based tools mainly focus on only a small subset of state or local designs.

To address these limitations, this study introduces a novel sampling method called scalable adaptive sampling for the creation of efficient and accurate Artificial Neural Network (ANN) models capable of considering a wide range of pavement design variables with nationwide applicability. 40 representative climate stations were selected across the contiguous US. The ANN models were trained using adaptively generated samples from Pavement ME simulations. The developed ANN models exhibited high accuracy in predicting fatigue damage and differential energy, enabling the estimation of long-term transverse cracking and joint faulting performance, respectively. The developed ANN-based M-E design tool offers flexibility for considering multiple variables during pavement designs. It matches the accuracy and surpasses the computational speed of the Pavement ME. By overcoming the limitations of the conventional ML-based M-E implementation, this research contributes to advancing efficient and reliable concrete pavement design practices.