Probabilistic Study of Liquefaction Response of Soil Based on Standard Penetration Test (SPT) Data Using Machine Learning Techniques

It is difficult and time consuming to evaluate the possibility for liquefaction using traditional experimental or empirical analysis techniques. This study offers SPT-based liquefaction datasets to evaluate the liquefaction potential of soil deposits using a variety of machine learning models: support vector regression (SVR), random forest (RF), and deep neural network (DNN). 300 datasets are generated to build machine learning models (210 random datasets for the training purpose and 90 random datasets for testing purpose). To predict the factor of safety (FOS) against liquefaction, these three machine learning models consider seven important input parameters, namely depth, total vertical stress, effective vertical stress, SPT blow count, fine content, earthquake magnitude, and peak horizontal ground acceleration. The relationship between soil and seismic factors, and the FOS against soil liquefaction was also investigated using a Pearson correlation matrix. The effectiveness of the machine learning models is evaluated using a variety of performance metrics, such as coefficient of determination (R2), Willmott's index of agreement, variance account factor, root mean square error (RMSE), mean absolute error and maximum absolute error. Based on performance criteria, the results indicate that DNN had the best prediction performance out of the three machine learning models. This refers to highest R2 = 0.972 and the lowest RMSE = 0.026 during the training phase, and maximum value of R2 = 0.870 and minimum value of RMSE = 0.062 during testing phase. The predictive power of the suggested model was also evaluated by the use of multiple performance indices, such as rank analysis, reliability index, regression curve, Taylor diagram, objective function criterion, and performance strength criterion. Additionally, these models' robustness was evaluated using external validation and comparative analysis.