From in-situ tests to probabilistic and sensitivity analysis using FEA and surrogate modelling

Abstract

This paper demonstrates a comprehensive workflow for probabilistic and sensitivity analysis of settlement of a hypothetical embankment case using realistic CPT data. The stratification reveals layers with different behaviour types, including clays, silt mixtures, sandy mixtures and sands. The dominant layer consists of soft clay which is susceptible to excessive deformation due to the embankment construction. Given the significant variations in the CPT readings, deformation analysis with deterministic parameters is insufficient. Instead, a series of finite element analyses were conducted with different material properties. The soft clay was modeled using the Soft Soil constitutive model, while other layers were modeled using the Hardening Soil Small Strain model in PLAXIS 2D. Model parameters were randomly selected from the parameter sets derived from individual CPT readings of the respective layer. In this way, each reading was considered as an independent measurement instance, with consistent derived parameters. Uniform sampling over depth maintained the actual distribution of the parameters within each layer. Additionally, the unit weight of the embankment was randomly chosen from its probability density function. Latin hypercube sampling was employed to ensure even sampling across the parameter space with a smaller number of samples. The resulting embankment settlements were processed in terms of the probability of exceeding a predefined maximum allowed value. To identify the significant contributing factors to settlement and deformation variation, Sobol’ global sensitivity analysis was performed. Uncertain parameters included representative features for each layer and the embankment's unit weight. The Sobol’ procedure requires considerably more simulations than required by the probabilistic analysis, which is prohibitive because of the computational demands. Therefore, several surrogate models were trained using machine and deep learning algorithms. To achieve acceptable performance, hyperparameters were tuned automatically over appropriate ranges. The best-scoring model was then employed to estimate the embankment displacements in the sensitivity analysis.