Evaluation of Liquefaction Potential Using Neural Networks Based on Adaptive Resonance Theory

Abstract

The incomprehensible loss of lives and extensive damages to transportation facilities caused by earthquakes emphasize the need for robust and reliable methods for evaluating the liquefaction potential of sites. Traditional methods for evaluating liquefaction potential are based on correlating data from the standard penetration test (blow count, N), cone penetration test (cone resistance, q_c), or the shear wave velocity (V_s) with the cyclic stress ratio. These methods are unable to incorporate the complex influence of various soil and in situ state parameters. This problem encouraged the development of numerous nontraditional methods such as artificial neural networks that try to learn and account for the influence of various soil and in situ state properties. The possibility of using neural networks based on adaptive resonance theory (ART) for the prediction of liquefaction potential was explored. These networks have been shown to be far more efficient and reliable than the commonly used backpropagation artificial neural network and other multilayer perceptrons. Two Fuzzy ARTMAP (FAM) models were developed and tested with q_c and V_s data obtained from past case histories. The q_c-and V_s-based FAM models gave overall successful prediction rates of 98% and 97%, respectively. The promising results obtained by the FAM models exemplify the potential of nontraditional computing methods for evaluating liquefaction potential.