A Comparative Analysis of Predicting Seismic Liquefaction Susceptibility of Dhaka Subway Project with a Machine Learning Approach.

Abstract

Seismic soil liquefaction is a dangerous phenomenon that occurs during seismic loading due to earthquakes. In this study, empirical formulas are used to assess liquefaction triggering based on the standard penetration test (SPT) data from the Dhaka Subway Project. After that three machine learning algorithms are applied to predict seismic liquefaction triggering of the obtained dataset. The first machine learning algorithm, Logistic regression, is a linear classification model. It implements the sigmoid function to generate binary outputs. The second machine learning algorithm is the Support Vector Machine (SVM) which represents supervised learning and is widely used as a classification and outliers detection algorithm. The third machine learning algorithm is the Artificial Neural Network (ANN) based on the Multi-layer Perceptron (MLP) theory, which uses a training algorithm called Levenberg-Marquardt backpropagation. Furthermore, this study also highlights the correlation between different soil parameters in triggering soil liquefaction. The developed models are then evaluated with confusion matrices which are later used to find out Overall Accuracy, Precision, Sensitivity, Recall (Specificity), F1 score, RMSE, and MAE. ROC curves are also used to evaluate these models and establish which model is the most effective.