Probabilistic Evaluation of Liquefaction Potential Using Multivariate Adaptive Regression Splines

Liquefaction poses major technical challenges for key infrastructures such as nuclear power stations and large earth dams. In recent years, several efforts have been made to assess the liquefaction potential of a site using simplified methods, namely the Seed and Idriss (1970) method, the Youd et al. (2001) method, the Idriss and Boulanger (2004) method, and the IS Code (IS 1893 (Part 1): 2016) method. In general, the problem of liquefaction is highly nonlinear in nature and the parameters involved, namely the engineering properties of soil and the earthquake characteristics, are all subjected to uncertainties. Although the role of soil plasticity in predicting the liquefaction potential in fine-grained soils is well recognized, none of the above-mentioned methods considered the effects of liquid limit and plasticity index in addition to the parameters such as N value, fines content, peak ground acceleration, and cyclic stress ratio. Keeping the above in view, the aim of this study is to (i) develop a comprehensive surrogate model considering all the above-mentioned six parameters using the multivariate adaptive regression splines (MARS), (ii) train and test the developed model using the dataset available in literature, and (iii) perform probabilistic analysis using the first-order reliability method (FORM) for predicting the liquefaction response of soils. Considering the computational efficiency, predictive accuracy, and the adaptivity associated with the developed model, the use of MARS in assessing liquefaction potential has been found to be promising.