Neural networks for fatigue cracking prediction using outputs from pavement mechanistic-empirical design

Fatigue cracking (FC) on asphalt pavements is one of the most critical performance indicators in the AASHTO Mechanistic-Empirical Pavement Design Guide (MEPDG). However, many recent studies have shown that the accuracy of the FC transfer function is unsatisfactory even after local calibration. This study approached this issue using the highly flexible neural networks through the deep learning framework called the Tensorflow. The data involved were extracted from the reports for the project NCHRP 01-37A, on which the calibration of the MEPDG transfer functions was based. The similarity between the FC transfer function and the neural network was discussed. A total of twelve neural network models divided into two groups were constructed. One group had only two nodes in the input layer and was used to represent the transfer function. The other contained an input layer of fourteen nodes that were used to introduce more relevant information for potentially higher predictive performance. The results indicated that, when properly designed, a neural network model could significantly outperform the FC transfer function. Comparing the FC transfer function with the network composed of an input layer of fourteen nodes and four hidden layers, the coefficient was increased from 0.008 to 0.613, while the mean absolute error dropped by 30%. Compared with the best model with an input layer of two nodes, the model with four hidden layers and fourteen nodes in the input layer increased the R 2 by 56% while decreased the mean absolute error (MAE) by 29%.

Automated summary

This study successfully constructed neural network models that outperform conventional FC transfer functions by using highly flexible neural networks in the Tensorflow framework.