Machine Learning Applications for Mechanistic-Empirical Concrete Pavement Design and AnalysisLi, Haoran (2023) Machine Learning Applications for Mechanistic-Empirical Concrete Pavement Design and Analysis. Doctoral Dissertation, University of Pittsburgh. (Unpublished) This is the latest version of this item.
AbstractThe American Association of State Highway and Transportation Officials Mechanistic-Empirical Pavement Design Guide (AASHTO M-E) enables the user to design pavement structures that meet long-term performance requirements with more economical solutions than traditional empirical design procedures. However, the prediction of pavement distresses using the AASHTOWare Pavement ME software is complicated and time-consuming. In addition, the limitation on the minimum concrete slab thickness leads to overly conservative designs for low-volume traffic. This study uses machine learning (ML) techniques to simplify and accelerate the M-E design for medium- to high-volume traffic roads and improve the M-E design for low-volume traffic roads. ML models were developed from Pavement ME simulations of fatigue damage and differential energy for the prediction of long-term performance, i.e., transverse cracking and joint faulting, respectively. A simplified M-E design procedure and a standalone alternative M-E design tool, named PittRigid ME, were developed to implement these ML models for the design and performance prediction of joint plain concrete pavements (JPCPs) for Pennsylvania conditions. PittRigid ME matches the Pavement ME performance predictions and designs at a fraction of the computation cost. To further expand the PittRigid framework for more design parameters, a new sampling method, scalable adaptive sampling, was developed. This sampling method has high scalability for developing ML models with a large number of input variables. Artificial Neural Networks (ANNs) were developed using Pavement ME-generated data for 40 climate stations located throughout the United States and a wide range of many categorical and continuous design variables. These ANN models match Pavement ME fatigue damage and transverse cracking predictions and enable computationally efficient implementation of the AASHTO M-E procedure for the design and analysis of concrete pavements. Multi-gene Genetic Programming (MGGP)-based ML models were developed for the optimal design of low-volume roads based on the AASHTO M-E framework. These models extrapolate Pavement ME simulations of pavement damage and performance for Portland cement concrete (PCC) slabs thinner than 150 mm (6 in.), which enables the use of thinner JPCPs for more rational and cost-effective design of low-volume roads. Share
Details
Available Versions of this Item
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|