1370887 - Machine Learning Solutions for Top-Down Cracking Design of Airport Rigid Pavement

The Federal Aviation Administration (FAA) rigid pavement design process is based on bottom-up cracking failure resulting from tensile stress at the bottom of a flat slab under aircraft loads. The FAA has a long-term goal to add top-down cracking failure mode to the FAARFIELD program. The existing design procedure is not suitable to support design for the top-down cracking failure mode. Critical stresses for rigid pavement design can be calculated by FEAFAA, the FAA three-dimensional finite element (3D-FE) program. However, direct use of 3D-FE methods in design software is typically far more time consuming than is acceptable. The objective of this research was to develop machine learning (ML) solutions to support design of airfield rigid pavements to resist top-down cracking. The ML model is intended to be used as a drop-in replacement for 3D-FE used by FAARFIELD to quickly calculate critical concrete stresses due to aircraft and thermal loads. The model targets rigid pavement design of airfields serving commercial aircraft heavier than 100,000 pounds gross weight. The ML model is a general model that supports individual dual (D), dual-tandem (2D) and-dual tridem (3D) gear configurations as well as a general model for full belly or landing gear configurations. The study proposed a conceptual design method based on cumulative damage factor (CDF) but suitable for top-down cracking design. The input required by the conceptual design method implies that the ML model needs to provide the distribution of transverse stress along the transverse joint and the distribution of longitudinal stress along the longitudinal joint. A synthetic database consisting of 127,000 input-output tuples was developed to train the ML model. The database contains distinct combinations of rigid pavement, thermal, and aircraft gear parameters that were input into FEAFAA 3.0 to determine the resulting stress distribution at the top of the slab. Researchers developed a new artificial neural network (ANN) method that predicts a dynamic functional evaluated over a continuous domain. The model is based on a modular deep learning method. The training operation was performed using backpropagation and the ADAHESSIAN numerical optimization algorithm. The models constructed with the new method are significantly more accurate than previous ML techniques for similar problems. The resulting model is a single model with one interface for all gear types in the data set. The ML model was compiled into a .NET-compatible library suitable for use in a program like FAARFIELD.