Seismically Induced Liquefaction Potential Assessment by Different Artificial Intelligence Procedures

Liquefaction triggering phenomenon during earthquake is one of the most complicated geotechnical problems due to the complex and heterogeneous nature of the soils. In this study, artificial intelligence based predictive models have been developed to assess the probability of liquefaction. In this context, the total 834 field data sets have been taken from the published literatures pertaining to past earthquakes to develop the machine learning models. Artificial intelligence-based regression techniques such as, Relevance Vector Machine (RVM), Genetic Programming (GP), and Multivariate Adaptive Regression Spline (MARS) have been utilized to estimate the liquefaction potential of a soil deposit. The relative efficiencies of the applied machine learning models have been ascertained by comparing various error indices such as Nash–Sutcliffe efficiency ( NS ), Weighted mean absolute percentage error ( WMAPE ), root mean square error (RMSE), Coefficient of determination ( R 2 ), variance account factor ( VAF ), mean absolute error ( MAE ), adjusted determination coefficient ( AdjR 2 ), Willmott’s Index of agreement ( WI ) etc. This study also suggests mathematical equations based on the obtained result for all models to compute the probability of liquefaction. The performance of proposed models was evaluated on the basis of various performance indices, actual versus predicted curve and rank analysis, etc. Additionally, Taylor diagram and regression error characteristic (REC) curve are also presented to check the effectiveness of the proposed models. On the basic of acquired results, it can be concluded that the GP model predicts the probability of liquefaction effectively compared to MARS and RVM models. This study shows the applicability of GP, RVM, and MARS models that offer a useful alternative tool for earthquake engineers to assess liquefaction conditions in liquefaction-prone areas.