Efficient Simulation Techniques for Reliability Analysis of Flexible Pavements Using the Mechanistic-Empirical Pavement Design Guide

Many sources of uncertainty are inherent in pavement design. These uncertainties must be incorporated systematically in a reliability analysis to compute their combined effects on the probability of failure of a given pavement structure. Monte Carlo simulation has been the technique of choice in the past to simulate the effects of uncertainties in input parameters on pavement distress and the resultant reliability analyses. The impractical computational time associated with a Monte Carlo scheme, however, has prompted the deferral of the implementation of similar techniques in the current Mechanistic-Empirical Pavement Design Guide (MEPDG). Instead, the reliability analysis implemented in the current MEPDG is performed on the basis of a simple assessment of the overall standard error of the predicted distress compared to the observed distress of the long-term pavement performance (LTPP) sections. It relies on a set of predetermined variability values obtained from a performance database instead of the site-specific input parameters that induce such uncertainty in distress predictions. Past efforts found that techniques (such as the Latin hypercube method) that require a substantially reduced number of simulations compared with Monte Carlo accuracy still suffered from the need for repeated execution of the MEDPG calculations. This study proposes to combine an efficient numerical scheme to conduct statistical simulations with the MEPDG calculations. It makes use of the concept of the representative linear elastic (LE) structure to minimize the number of repeated executions involved in simulations. The numerical scheme can be combined with any simulation technique of random variables to perform a reliability analysis of flexible pavements. The relative merits of Monte Carlo simulation, Latin hypercube simulation, and Rosenblueth’s $2K+1$ point-estimate method are compared. The simulations show that the Latin hypercube method is an efficient alternative to the computationally intensive Monte Carlo technique. On the other hand, although Rosenblueth’s $2K+1$ point-estimate method is much simpler, it is not capable of capturing the important attributes of the distribution of either input or output variables.