Evaluation of Liquefaction Potential of Soil Using Soft Computing Techniques

This study examines the potential of soft computing techniques, such as multi-adaptive regression spline (MARS), Bayesian regularization neural network (BRNN), Levenberg–Marquardt neural network (LMNN) vis a vis statistical regression technique, like multiple linear regression (MLR)-based classification approaches to evaluate liquefaction potential of soil in terms of liquefaction index (LI) from a large database consisting of post liquefaction SPT measurements and liquefaction field observations. The liquefaction classification accuracy: 94.44% (LMNN) and 94.44% (MARS) of the developed LI models is found to be better than that of available artificial neural network (LMNN) model (88.37%), support vector machine model (94.19%) and multi-gene genetic programming (MGGP) model (94.19%) on the basis of the testing data. A ranking system is used to evaluate the above models based on different statistical performance criteria like correlation coefficient (R), Nash–Sutcliff coefficient of efficiency (E), log normal probability distribution of ratio of predicted LI (LIp) to observed LI (LIm) etc. Based on the above ranking criteria LMNN model is found to be better than BRNN, MGGP, MARS and MLR models. Model equations based on the above techniques are also presented for geotechnical engineering professionals.