Artificial Neural Networks Applied to the Seismic Design of Deep Tunnels

Abstract eng:
Seismic response of underground structures is controlled by the earthquake-induced ground strain field and its interaction with the structure. Focusing on the response of the cross section, for simple geometry several closed-form solutions are available in the technical literature to compute the earthquake-induced stress increment in the lining. All these solutions are functions of the shear strain field which is the cause of the ovaling of the cross section. Since no direct measure of the transient ground strain during earthquake are available, it is common practice to indirectly compute the peak ground strains through simplified formulas based on simple assumptions of plane wave propagation in a homogeneous medium. A careful review of the damages caused by the earthquakes to underground structures shows that most damaged tunnels were located in the vicinity of the causative fault. Under such conditions ground motion is affected by near-fault effects and the induced strain field is quite complex. Therefore the use of simplified formulas may lead to a severe underestimation of the ground maximum strain. The strain field at depth can be evaluated numerically through the computation of synthetic time histories, however this procedure is rather involved, time consuming and requires numerous seismological input parameters. This paper illustrates the result of a numerical study in which an Artificial Neural Network (ANN) has been trained to predict the shear strain field in a neighbourhood of a seismogenic fault. The strain field was computed through numerical differentiation of synthetic displacement time histories obtained using the extended kinematic source model by Hisada and Bielak (2003). The reactivation of a fault located in the Sannio region (Southern Apennines, Italy) has been selected as a case study, since the fault is placed in the vicinity of an existing deep rock tunnel which is part of an important railway line in Southern Italy. The training of the ANN was conducted for a seismic source with varying magnitude, geometry and focal mechanism. Observation points at different strike and depth from the ground surface were considered. The computed results show the capability of ANN to predict the earthquake-induced strain field at depth in near-fault conditions and for varying seismological parameters.

• Export as: BibTeX | MARC | MARCXML | DC | EndNote | NLM | RefWorks
• Add to your basket