Empirical model for liquefaction resistance of soils based on artificial neural network learning of case histories

In recent years, the method of artificial neural network (ANN) has been applied to solving geotechnical problems, especially for mapping high complexity soil response with a multi-parameter input vector. Successful applications of simple but robust, feed-forward neural networks have been reported. In this paper, a procedure for developing an empirical model for assessing liquefaction potential of soils based on ANN learning of case histories is presented. The intended model is based on piezocone penetration test (CPTu). A pore pressure parameter B_q (excess pore pressure ratio) is included in the formulation of soil behavior index I_c, which, in turn, is used along with the normalized cone tip resistance q_c1N to characterize liquefaction resistance. For ANN learning, each case history consists of three input variables, q_c1N, I_c, and CSR (cyclic stress ratio, as a proxy to liquefaction loading), and one output variable, binary field observation (yes or no), and the learning of case histories is carried out with a feed-forward back-propagation neural network. The successfully learned or trained neural network is then used to search for a limit state for liquefaction triggering, and the searched data is employed to develop the intended empirical model for liquefaction resistance. Characteristics and applications of the developed ANN-based empirical model are presented. Copyright ASCE 2008.