Estimation and Validation of Gaussian Process Surrogate Models for MEPDG-Based Sensitivity Analysis and Design Optimization

摘要

The Mechanistic-Empirical Pavement Design Guide (MEPDG) is a powerful predictor ofpavement distress, but it is computationally expensive to evaluate. Analyses requiring largenumbers of MEPDG evaluations, such as sensitivity analysis and design optimization, becomeimpractical due to the computational expense. These applications are important in achievingrobust, reliable, and cost-effective pavement designs. This paper develops Gaussian processsurrogate models that, with a trivial amount of computational expense, accurately approximatethe results of the MEPDG for each relevant distress mode. The GP is validated according to threemodel validation metrics: average predictive percent error, predictive coefficient ofdetermination, and Bayes factor. The GP models are then exploited for purposes of sensitivityanalysis and design optimization, making these tasks computationally affordable.