Abstract

The Federal Aviation Administration (FAA) Airport Pavement Technology Research and Development Office initiated a research project aimed at incorporating top-down cracking failure mode into the FAA Rigid and Flexible Iterative Elastic Layered Design (FAARFIELD) program. This pavement design methodology uses an iterative process of damage accumulation, requiring accurate stresses calculations under aircraft and environmental loads. FAARFIELD employs a three-dimensional finite element (3D-FE) model to compute these critical stresses. While the 3D-FE is considered the industry standard for accurately predicting pavement stresses, the execution time of a single simulation required for a top-down cracking design can extend up to 30 minutes. Given that the iterative design process involves numerous simulations for each aircraft, the overall computational time becomes prohibitively expensive. To address this challenge, the FAA developed a machine learning (ML)-based stress prediction model to replace the 3D-FE model. As originally developed, the ML model was designed specifically for four-layer airfield rigid pavements serving commercial aircraft heavier than 100,000 pounds.

The objective of this research was to enhance the applicability of the previously developed ML model to include relatively thin rigid pavements at facilities serving light-load aircraft, such as general aviation airports, and three-layer rigid pavements designed to support aircraft heavier than 100,000 pounds. A comprehensive database was created for training the ML model that consisted of the results of 250,000 3D-FE simulations encompassing a wide range of rigid pavement material and thickness parameters, thermal loads, and aircraft configurations. This database was combined with the existing database for the four-layer pavements. The database supports a general model that includes a wide range of weight and configurations of single and full landing gear.

Researchers developed a new artificial neural network (ANN) method that predicts a dynamic function for a continuous waveform prediction. The training operation was performed using backpropagation and the ADAHESSIAN numerical optimization algorithm. The model rapidly estimates stress distribution along slab edges resulting from aircraft and thermal loads. The prediction error across all models is less than 3.9 pounds per square inch (psi), resulting in less than 0.5 inch of error in the slab design thickness. The ML model was compiled into a .NET-compatible library suitable for use in FAARFIELD.