Design and analysis of computer experiments for numerical modeling of mechanized tunneling-induced settlements using surface monitoring data

Pressurized-face TBMs, such as EPB machines, are designed to limit ground movements, but excavation still induces deformations that propagate to the surface as settlements, potentially affecting structures. Controlling these settlements relies on operational parameters like face pressure, grouting pressure, and advance rate. However, their effects are difficult to predict due to geotechnical uncertainties and the inherently nonlinear and three-dimensional nature of the process. This creates a need for predictive tools that link operational settings to induced settlements while accounting for uncertainty.In practice, empirical, analytical, or two-dimensional numerical models are commonly used because they are simple and fast. Yet, they rely on strong assumptions, notably plane strain, and fail to capture the full three-dimensional excavation process or explicitly integrate operational parameters. Three-dimensional numerical models overcome these limitations by providing a more realistic representation of the physics and TBM operations. Their major drawback is computational cost, which restricts their use for rapid decision-making and large-scale uncertainty analysis.To address this, the thesis proposes a methodology based on the Design and Analysis of Computer Experiments (DACE) framework. A parametric 3D finite element simulator is first developed to model the excavation process and extract settlement-related quantities of interest. To reduce computational cost, an Accuracy-Cost Model Reduction (ACMR) strategy is introduced to quantify discretization and boundary effects and define optimized simulation settings.Gaussian process metamodels are then constructed to approximate the simulator at low cost. These metamodels enable global sensitivity analysis to identify the most influential parameters, guiding model calibration. A key contribution is the development of a data processing pipeline that transforms noisy monitoring measurements into consistent, dynamic quantities of interest, ensuring comparability with simulation outputs.Calibration is formulated as a Bayesian set-inversion problem, aiming to identify all input configurations consistent with the observations rather than a single solution. This is solved using an adaptation of Bayesian subset simulation (BSS), combining Gaussian process metamodeling and sequential Monte Carlo methods. Dedicated diagnostics are introduced to distinguish uncertainties arising from the metamodel and from the sampling procedure, and to guide adaptive refinement.Applied to a real case study, the framework reduces computational cost by a factor of 5 to 10, identifies the dominant geotechnical and operational parameters, and determines sets of input configurations consistent with observed settlements. Results also highlight the need for additional simulations to refine narrower target regions.