Shear wave Velocity-Based Machine Learning Modeling for Prediction of Liquefaction Potential of Soil

A comparison of efficacies of four machine learning methods such as multiadaptive regression spline (MARS), Bayesian regularization neural network (BRNN), Levenberg-Marquardt neural network (LMNN), and multiple linear regression (MLR) in generating mathematical equations for evaluating liquefaction potential of soil is presented in this paper. These models are generated using a comprehensive dataset comprising of post-liquefaction shear wave velocity measurements and field manifestations of liquefaction occurrences. The correct rate of prediction of liquefaction and non-liquefaction events on test data by the developed models: MARS (90.32%) and LMNN (87.10%), was found to be highly competitive with that of the available extreme learning machine (ELM) (88%) and multigene genetic programming (MGGP) (86%) models, respectively. Further, a ranking method is utilized to assess the developed models as well as the existing ELM and MGGP models based on several statistical performance measures. According to this ranking analysis, the LMNN model outperformed the MGGP, ELM, MARS, MLR, and BRNN models. The sensitivity analysis revealed that the cyclic stress ratio (CSR7.5) was the most influential parameter, followed by effective vertical stress (sigma'v), fines content index (FCI), and shear wave velocity (Vs). This investigation also presents the model equations for use by practicing engineers.