Parameter influence analysis in a 3D TBM model via sensitivity analysis and Gaussian process metamodels

Urban tunnel excavation with tunnel boring machines induces ground movements that can affect nearby structures. Three-dimensional finite element models (FEM) are widely used to predict these settlements, but their high computational cost limits direct exploration of parameter influence. This work presents a 3D FEM simulator of mechanized tunneling and a methodology to quantify the impact of both numerical and physical inputs on settlement predictions. First, an accuracy-cost model reduction study evaluates the effect of domain dimensions and mesh densities on a small number of scalar quantities of interest extracted from simulated settlement fields. Empirical error models are fitted and used to select a reduced configuration that balances accuracy and runtime. Second, Gaussian process models are trained on simulation data from the reduced configuration and validated using exact leave-one-out cross-validation. These metamodels enable the computation of Sobol’ sensitivity indices with quantified uncertainty, identifying the most influential geological, operational, and loading parameters. The proposed framework reduces the cost of sensitivity analysis for computationally intensive 3D tunneling simulations, supporting input screening and dimensionality reduction for design and calibration.