A Comparative Study of Soil Liquefaction Assessment Using Machine Learning Models

Abstract

Liquefaction of saturated granular soils is marked by the total loss of shear strength of soil under dynamic cyclic or transient loading conditions due to excess pore water pressure that builds up to produce a soil regime that mechanically performs as a liquid. The cone penetration test (CPT) is widely recognized as a means of evaluating liquefaction susceptibility. This study presents a comparative supervised machine learning-based assessment for CPT-based liquefaction data. In particular, this study views soil liquefaction as a binary classification problem, whether the soil is liquefied or not, by utilizing three supervised machine learning classifiers: support vector machine, Decision Trees, and Quadratic Discrimination Analysis. To build the supervised machine learning models, three different soil characterization data sets were selected by performing CPTs at specific locations. The first input data (input data-1) is constructed as a function of the Mean Grain Size (D₅₀), Measured CPT Tip Resistance (q_c), Earthquake Magnitude (M), and Cyclic Shear Resistance (CSR). The second input data (input data-2) employed D₅₀, Normalized CPT Tip Resistance (q_c−1), M, CSR. Finally, the third input data (input data-3) consists of D₅₀, q_c−1, M, the Maximum Ground Acceleration (a_max), Effective Vertical Overburden Stress, and Total Overburden Stress. The significance feature analysis shows the most important feature for assessing liquefaction susceptibility in the soil using input data for model 1 is measured CPT Tip Resistance, for input data model 2 it is normalized CPT Tip Resistance, and finally, for input data model 3, it is measured CPT Tip Resistance. Conclusively, this study proposed simple and quick approaches to evaluate soil liquefaction susceptibility without complicated calculations.

Appendix

This section describes the machine learning methods used to predict a new liquefaction model. The liquefaction potential was evaluated using CPT databases. The next parts present a concise explanation of the basic knowledge of these methods. The validation and quantification were performed by using overall accuracy, precision.

1.1 Supervised machine learning classifiers

This section summarizes some of the supervised machine learning techniques used in this article. In particular, the following methods are reviewed: Decision Tree, Support Vector Machine (SVM), and QDA.

1.1.1 Decision tree

The decision tree is one of the most popular supervised machine learning classifiers. This classifier operates by dividing the feature space into axis-parallel rectangles and labeling each rectangle with one of the two classes (Yang et al. 2018).

1.1.2 Support vector machine (SVM)

Support Vector Machine is a supervised machine-learning method that has been used for classification and regression analysis. The linear vector machine algorithm takes as input data consists of training examples $\left({\text{x}}_{1} .{\text{y}}_{1}\right)$…$\left({\text{x}}_{\text{N}} .{\text{y}}_{\text{N}}\right)$ where the points ${\left\{{\text{x}}_{\text{i}}\right\}}_{{{\rm i}=1}}^{\text{N}}\subseteq{\text{R}}^{\text{P}}$ and the labels ${\text{y}}_{\text{i}}\in\text{Y} = \left\{{\pm 1 }\right\}$ for every ${\text{i}}$ and return a $\mathrm{p}-1$ hypersurface (Huang et al. 2018).

1.1.3 Quadratic discriminant analysis (QDA)

In QDA, the decision boundary is assumed to be a quadratic surface. In other words, a QDA classifier tries to find a quadratic surface that best separates the training set data. From this understanding, QDA can be considered as a generalization of linear classifiers (Ghojogh and Crowley 2019).