Metamodel-Based Prediction of Structural Damages due to Tunneling-Induced Settlements

Abstract

Tunneling-induced settlements in soft soil are inevitable and often suspected to impair buildings along the alignment. Public focus is thus targeted toward the consequences rather than the benefits of tunneling projects. To reduce the risk of damage, costly compensation is frequently applied in practice when crossing below critical structures, and critical scenarios are usually assessed via deterministic analyses of analytic models. These models are well known and overestimate potential damage since they conceive of buildings as idealized, simple Timoshenko beams without considering soil–structure interaction. The present study introduces an approach to providing an efficient prediction of realistically idealized structures prone to settlement-induced damages based on random input. This approach uses a nonlinear finite-element simulation of façades to render damage assessment. Global sensitivity analysis that uses both elementary effects and Sobol indices identifies random but irrelevant input that is consequently fixed to mean values. Response surfaces yield surrogate models for subsequent damage prediction via polynomial regression. As input, these surfaces employ ordinary distribution functions and extreme value distributions to ensure predictability at the tails. The inherent error of any prediction with response surfaces is first reduced by cross-validation and then comparatively rated with alternative forecasts through artificial neural networks. Finally, the mean probabilities of occurrence and standard deviations are determined considering the imperfection of the surrogate models for the classified damage events.