Artificial neural networks applied to the seismic design of deep tunnels

Student: Teraphan Ornthammarath
Supervisors: Dr. Carlo G. Lai, Dr. Mirko Corigliano

ABSTRACT

The underground structure responses in competent rock are widely accepted to conform to the surrounding ground during the earthquake which is different from the aboveground structure responses. Additionally, past study suggested that underground structures in general were less severely affected than surface structures at the same geographic location. However, in the 1995 Kobe, 1999 Chi-Chi, and 1999 Kocaeli earthquakes, the damages of underground structures in these events show that most tunnels were located in the vicinity of the causative fault. One of the main contributions of these damages is the near-fault effect. From the past observations, the near-field ground motions produce ground motion characteristic in the vicinity (<10-25 km) different from that in the far-field because of the directivity and fling step effects. It is important from the practical design point of views to evaluate the seismic performance of underground structures at a particular site, especially in near field. This study presents a simplified method to predict the maximum shear strains around the fault by using Artificial Neural Networks (ANNs). Since the deformation of underground structures, both longitudinal and transversal, is mainly caused by the longitudinal and shear strains respectively in terms of the whole cross section, the proposed method is then based on identification of these shear strains by ANNs.

The proposed method is applied to the "Ariano Irpino" fault located in Southern Italy that was subjected to the December 5, 1456 earthquake. The near-field ground motion model developed by Hisada and Bielak [2003] had been performed as this fault with assumed ground profile in that area. The observation point is the point where seismometers or accelerograms would be place to record the ground motion characteristics. The observation points had been assumed to be laid next to the fault in different directions. For this study, it was assumed that we have observation points only in 100- and 600-meter depths. These synthetic data would be used as a training data for ANNs to learn the near-field ground characteristics. From this assumption, the trained ANNs would be able to predict the maximum shear strains in other different directions and depths. The computed results show that the ANNs has a possible capability to predict the maximum shear strains around the fault vicinity.

You may download a digital version of this MSc dissertation here.