Assessment of seismic site effects in 2-D alluvial valleys using neural networks

There is an increasing need to fill the gap between research and engineering practice as far as consideration of complex seismic site effects is concerned. To address this problem in an innovative way, we have developed a pattern recognition approach based on neural networks as a discriminant tool. The classification algorithm has been trained with the results of a large number of numerical simulations of the seismic response of 2-D layered valleys with both symmetric and asymmetric shape, using as excitation several real accelerograms. The different phases of the classification analysis are explored, namely: a) definition of the set of features (i.e., pattern attributes); b) definition of appropriate classes for discrimination of hazard levels; c) training of the neural network; d) interpretation of results using a structured neural network architecture and evaluation of the most critical features that govern the classification; e) validation of the trained network on real case histories. Among the most interesting conclusions of this study, is that the most relevant information for predicting the site-related hazard in a 2-D valley configuration is condensed in the local 1-D resonance frequency and the global 2-D resonance of the valley (f01D and f02D). If one considers as features only f01D and f02D the performance of the classification tool is only slightly lower than that obtained using a larger set of features.