Comparison Between Two Surrogate Models for Embankment Earthquake-Liquefaction-Induced Settlements Prediction.

Abstract

The development of quick, easy-to-handle, surrogate models of complex simulations is a key issue to derive a rapid approximation of more expensive models. For instance, Lopez-Caballero and Khalil (2018), performed multiple finite element (FE) numerical simulations to assess the effect of the liquefaction-induced settlements of the soil foundation of an embankment due to real earthquakes. However, due to computational limitations of the FE analysis, they proposed a Gaussian Process (GP) emulator to represent the output of the more expensive FE model. Furthermore, they derived the analytical fragility curves constructed on the basis of the output of the nonlinear dynamic surrogate model. The aim of this work is to confront two different kinds of meta-models techniques, based on the same earthquake-liquefaction-induced settlements data. Hence, we proposed a comparison between two different surrogate models: (1) the GP model proposed by the authors and (2) an Artificial Neural Network (ANN) model proposed under the scope of this work. A comparison between the resultant fragility curves of the levee using both surrogate models is discussed, together with the impact of both meta-models in terms of fragility curves and its corresponding uncertainty. Finally, the main advantages and drawbacks of the surrogate models are highlighted.