Alternative Approaches to Determining Robust ANN Based Models for Predicting Critical Airport Rigid Pavement Responses

A research version of the FAA Rigid and Flexible Iterative Elastic Layer Design (FAARFIELD) Program (version 2.0) was developed; it includes top-down cracking along with bottom-up cracking as failure modes to support calculation of more reasonable pavement thicknesses than previous versions. The software employs three-dimensional finite element (3D-FE) response models that take a significant amount of analysis time for computing critical pavement responses. This paper proposes two different approaches for implementing an artificial neural networks (ANN) based prediction model for more efficiently computing critical rigid pavement responses for a given aircraft type: (1) ANN models trained several times for each case with determination of a final ANN model based on the maximum correlation coefficient (R²), and (2) an approach using the average of network outputs after several training activities. A total of 1,000 cases, 500 for mechanical loading and 500 for simultaneous mechanical and thermal loading, were simulated in batch mode by varying different input parameters under a Boeing B767-300ER airplane loading. It was found out that the final ANN models produced similarly accurate critical pavement responses for both approaches so that one approach could not be considered superior to the other. However, the study highlights the importance of training the ANN models several times using a random initialization of weights before arriving at the best-performance model.