RESEARCH ARTICLE Liquefaction prediction using support vector machine model based on cone penetration data Pijush SAMUI * Centre for Disaster Mitigation and Management, VIT University, Vellore-632014, India * Corresponding author. E-mail: pijush.phd@gmail.com © Higher Education Press and Springer-Verlag Berlin Heidelberg 2013 ABSTRACT A support vector machine (SVM) model has been developed for the prediction of liquefaction susceptibility as a classification problem, which is an imperative task in earthquake engineering. This paper examines the potential of SVM model in prediction of liquefaction using actual field cone penetration test (CPT) data from the 1999 Chi-Chi, Taiwan earthquake. The SVM, a novel learning machine based on statistical theory, uses structural risk minimization (SRM) induction principle to minimize the error. Using cone resistance (q c ) and cyclic stress ratio (CSR), model has been developed for prediction of liquefaction using SVM. Further an attempt has been made to simplify the model, requiring only two parameters (q c and maximum horizontal acceleration a max ), for prediction of liquefaction. Further, developed SVM model has been applied to different case histories available globally and the results obtained confirm the capability of SVM model. For Chi-Chi earthquake, the model predicts with accuracy of 100%, and in the case of global data, SVM model predicts with accuracy of 89%. The effect of capacity factor (C) on number of support vector and model accuracy has also been investigated. The study shows that SVM can be used as a practical tool for prediction of liquefaction potential, based on field CPT data. KEYWORDS earthquake, cone penetration test, liquefaction, support vector machine (SVM), prediction 1 Introduction Liquefaction of saturated sandy soils during earthquakes causes building settlement or tipping, sand blows, lateral spreading, ground cracks, landslides, dam instability, high embankment failures and other hazards. Prediction of liquefaction of saturated sandy soils due to an earthquake is an important task in earthquake geotechnical engineering. Since it is very difficult to get high-quality undisturbed samples of sandy soils, in situ tests have been used to determine the liquefaction resistance of saturated sandy soils. The method of liquefaction resistance based on standard penetration test (SPT) data has been developed by seed and his coworkers [1,2]. However, there are several limitations in using their methodology to determine the liquefaction resistance of saturated sandy soils [3,4]. Because of its reliability, speed, economy and continuity of profiling, the CPT test is considered a superior technique for determination of liquefaction resistance [3,5,6]. Liquefaction analysis based on probabilistic and statistical methods have been done by many researchers [7–10]. But all of the above methods have been developed based on some empirical formulae, which are associated with some inherent uncertainties. More recently artificial neural network (ANN) model have been used for prediction of liquefaction potential as a classification problem [11–14]. Although this is successful in many regards, a major disadvantage, compared to other statistical models is that they provide no information about the relative importance of the various parameters involved, as also implied by some previous studies [15]. It has also been noted that as the knowledge acquired during training is stored in an implicit manner in the ANN, it is very difficult to come up with a reasonable interpretation of the overall structure of Article history: Received Oct. 10, 2012; Accepted Nov. 26, 2012 Front. Struct. Civ. Eng. 2013, 7(1): 72–82 DOI 10.1007/s11709-013-0185-y