Surrogate-assisted uncertainty modeling of embankment settlement

Abstract

The structural optimization of basal reinforced piled embankments is usually conducted by examining design alternatives while ignoring the inherent variability of soil properties and studying only a limited number of structural variables. As an alternative, this paper proposes a hybrid modeling framework to introduce soil property uncertainty into embankment settlement calculations. This is important because settlement is critical in the serviceability assessments considered during structural optimization. The proposed framework consists of uncertainty modeling, finite element method, surrogate modeling, and probabilistic analysis. More specifically, a neural network with Monte Carlo dropout that accounts for uncertainty is employed to correlate the soil properties which affect the long-term performance of embankments over soft clay. Next, a coupled finite element analysis is performed using two constitutive soil parameters generated by the neural network to predict post-construction settlements. Combining the finite element (input source) with a surrogate model (data-driven approximation) yields substantial settlement outcomes for structure evaluations. A case study is then used to validate the effectiveness and applicability of this framework. Finally, an exhaustive search approach is used to design a cost-effective improved ground within ultimate and serviceability limit state constraints. Pareto front is computed using a logistic function at different settlement reliability levels.