Study of MEPDG Sensitivity Using Nonparametric Regression Procedures: Fid-Bau Portal

Study of MEPDG Sensitivity Using Nonparametric Regression Procedures

Tarefder, Rafiqul A. / Sumee, Nasrin / Storlie, Curtis

Because the new Mechanistic-Empirical Pavement Design Guide (MEPDG) includes numerous inputs, a sensitivity analysis using the Monte Carlo approach is not practical because it requires thousands of MEPDG runs. Instead, nonparametric regression procedures can be very useful to perform MEPDG sensitivity analysis. In this study, nonparametric regression procedures such as multivariate adaptive regression splines and gradient boosting machine are employed to identify inputs that contribute significantly to the outputs. Thirty inputs are used to randomly generate 750 input combinations by using Latin hypercube sampling. Using four-layer pavement geometry [two asphalt concrete (AC), base, and subgrade layers], simulations are run in MEPDG software to produce a time series of predicted distresses such as roughness, rutting, and cracking. Sensitivity analysis resulted in three groups of inputs to which the output is (1) highly sensitive, (2) moderately sensitive, and (3) minimally sensitive. Results show that roughness is highly sensitive to traffic input variables such as annual average daily truck traffic (AADTT), percentage of trucks in the design lane, and thickness of bottom AC layer. AC rutting is highly affected by AADTT, percentage of trucks in design direction, and tire pressure. Three major factors for total rutting, longitudinal cracking, and alligator cracking are AADTT, percentage of trucks in design direction, and thickness of bottom AC layer. In addition to these, alligator cracking is highly sensitive to percentage of air voids in the bottom AC layer. Transverse cracking is highly sensitive to the percentage of trucks in the design lane, percentage of Class 11 vehicles, plastic limit, thickness of base layer, effective binder content of top AC layer, and climate. Among all of the inputs, the thickness of the AC layer is highly interactive with other input variables.