Application of cross validation approaches to examine soft computing systems for forecasting soil liquefaction

The objective of this research is to assess the liquefaction potential of typical penetration test data using five distinct machine learning models: linear regression (LR), random forest (RF), decision tree (DT), stochastic gradient descent (SGD), and multilayer perceptron (MLP). Additionally, three types of divisional approaches are employed: Grid Search cross-validation (Gridsearch CV), the k-fold divisional approach, and Fuzzy Clustering Means (FCM). The models are constructed using seven input variables: depth of soil strata, corrected penetration blow number, total vertical stress, fineness content, maximum horizontal acceleration, total effective stress, and groundwater table (GWT). The input parameters are derived from the deterministic approach developed by Idriss and Boulanger (2006) for determining liquefaction potential. Several statistical parameters are evaluated to determine the liquefaction forecasting capabilities of the selected models. Out of all the developed models, RF by the k-fold technique has the best predicting accuracy with (R²=0.874, R=0.935, and RMSE=0.108) for the testing phase and (R²=0.90, R=0.95, and RMSE=0.048) for the training phase when considering only the statistical features. Finding the optimal model becomes more difficult when testing and training phases are taken into account. Score analysis is thus carried out for this investigation. The RF model by k-fold cross-validation shows the highest score of 80, indicating the most effective model. The provided models could correctly predict the potential of liquefaction in intricate real-world scenarios and technological difficulties associated with earthquakes.