A data-driven approach using surrogate models and non-deterministic optimization techniques for calibration of soil parameters and sensitivity analysis: application to a rockfill dam
- Authors
- Publication Date
- Aug 07, 2023
- Source
- Espace ÉTS
- Keywords
- Language
- English
- License
- Unknown
- External links
Abstract
The development of advanced numerical models for designing and assessing the safety of complex structures such as rockfill dams heavily relies on the availability of significant computational resources. The intricate structure of rockfill dams, which consists of various zones with varying soil parameters, makes the models highly uncertain. Minor variations in some soil parameters can significantly impact the expected behaviour of the structure, making it challenging to determine the geomechanical parameters required for effective modelling. However, laboratory or in situ tests and empirical relationships from the literature are the general approaches to estimating these parameters. Nonetheless, these measures do not accurately depict the insitu characteristics of the dam. In this context, uncertainty and global sensitivity analysis have been carried out to determine the influence of constitutive soil parameters on the behaviour of a rockfill dam. In parametric studies, surrogate models are helpful in approximating the relationship between inputs (soil parameters) and outputs (displacement) and thus effectively reduce computational costs. Surrogate models, built using methods such as polynomial chaos expansion and deep neural networks, calculate the Sobol indices required for identifying the impact of soil parameters on dam behaviour. Two parameters, Shear modulus and specific weights, are considered more sensitive input random variables from which uncertainties occur. This thesis proposes a highly effective data-driven approach that utilizes deep neural networks and optimization algorithms to estimate in situ values of soil parameters for a rockfill dam located in Quebec. Extensive analysis of inclinometer displacement measurements was carried out using a 2D finite element model, with the computational domain being meticulously divided into subdomains to account for the variability of material properties. In order to expedite computations, surrogate models were employed in lieu of the complete FEM model. To solve the minimization problem, stochastic optimization algorithms such as Genetic algorithm (GA), Particle Swarm Optimization (PSO), and Differential evolution (DE) were thoroughly evaluated and compared. Another contribution of this thesis is to present a novel technique for enhancing safety and stability analyses by identifying parameters through a hybrid optimization method. A deep neural network surrogate model is established and a multi-objective function formulation is used to weigh the difference between predicted and actual displacements. The soil parameters are identified using a hybrid Particle Swarm-Genetic Algorithm applied to data from four inclinometers installed in two different cross-sections of the dam. The study compares the effectiveness of the Mohr-Coulomb (MC) and Hardening Soil (HS) models, showing that the HS model provides the closest values to the measured onsite data. The research concludes by presenting the optimal soil parameters for the Romaine-2 dam and highlights the effectiveness of DNNs and hybrid optimization in solving inverse problems.