Liquefaction refers to the phenomenon in which soil loses its shear strength due to full saturation, particularly in loosely packed granular soils such as silt or sand, causing the soil to behave similarly to a
liquid. The Cone Penetration Test (CPT) is a widely adopted method for evaluating liquefaction potential, as it provides reliable estimations of the mechanical parameters of sands.
A study conducted by Hanandeh et al. from the Department of Civil Engineering at Al-Balqa Applied University, Salt, Jordan, and published in Geotechnical and Geological Engineering in June 2022, undertook a comparative analysis by developing three supervised machine learning models: Decision Tree (DT), Support Vector Machine (SVM), and Quadratic Discriminant Analysis (QDA) for assessing liquefaction susceptibility. Each model maps the occurrence of liquefaction to a specific set of parameters.
The study utilized a dataset comprising 94 samples, with 53 classified as liquefiable and 41 as non-liquefiable, divided in a 70:30 ratio for training and testing purposes. The soil types in these locations ranged from silty sand to sandy silt, with measured depths from the CPT test varying between 1.3 and 15.1 meters. The tip resistance (qc) values ranged from 0.38 to 20.6 MPa, while the measured total stress varied from 31.4 to 290.3 kPa, and effective stress ranged between 13.9 and 227.5 kPa. Peak ground horizontal acceleration at the ground surface varied from 0.15g to 0.5g.
The authors developed machine learning models based on three datasets characterized by selecting CPT performance at specific locations. The first dataset included input parameters such as CPT tip resistance, cyclic shear resistance, mean grain size, and earthquake magnitude. The second dataset introduced normalized CPT tip resistance along with the other parameters. The third dataset encompassed maximum ground acceleration, earthquake magnitude, mean grain size, normalized CPT tip resistance, total overburden stress, and effective vertical overburden stress. The output from the models was categorized as either “happens” (1) or “non-happens” (0) regarding liquefaction occurrence, thus framing it as a classification problem.
Analysis revealed the following accuracies: for the first dataset, SVM achieved 0.85, QDA reached 0.94, and DT attained 0.71. In the second dataset, SVM recorded an accuracy of 0.93, QDA 0.90, and DT 0.71. For the third dataset, the accuracies were SVM at 0.87, QDA at 0.91, and DT at 0.92. Consequently, in the first dataset, QDA demonstrated superior performance, while SVM excelled in the second dataset, and DT performed best in the third dataset.
Furthermore, the authors conducted feature importance and model sensitivity analyses. Feature importance assesses how sensitive a classifier is to changes in specific features. In the first dataset, QDA was the best-performing model, with feature importance ranked in descending order as follows: tip resistance, cyclic stress ratio, earthquake magnitude, and mean grain size. In the second dataset, SVM outperformed the other models with features ranked as: normalized CPT tip resistance, cyclic stress ratio, earthquake magnitude, and mean grain size. For the third dataset, DT outperformed with features ranked as: mean grain size, total effective overburden stress, effective overburden stress, earthquake magnitude, measured CPT tip resistance, and peak ground acceleration at the ground surface.
The findings indicate that in dataset 1, four input variables were identified, representing the Cyclic Stress Ratio (CSR), mean grain size (D50), earthquake magnitude (M), and measured CPT tip resistance. Dataset 2 similarly included mean grain size, earthquake magnitude, CPT tip resistance, and CSR, with the addition of normalized cone tip resistance as a new variable. The analysis demonstrated that the predicted values of liquefaction closely aligned with the observed liquefaction in the proposed models.
Liquefaction in saturated sand is an important topic in geotechnical design. The CPT method has proven effective for soil exploration and analysis. This study employs machine learning to estimate liquefaction
occurrence using CPT data, proposing three models with different methods. Dataset-1, using QDA, outperformed the others, achieving a recall score of 1, accuracy of 0.97, and precision of 0.94. Dataset-2, based on SVM, had a recall score of 0.90, while Dataset-3, using a decision tree, achieved a recall score of 0.95. While recall alone was insufficient for determining the best classifier, additional metrics like precision and accuracy were utilized for final evaluation. The findings demonstrate that machine learning methods are effective tools for predicting liquefaction without relying on traditional calculation methods.
Reference
Hanandeh, S. M., Al-Bodour, W. A., & Hajij, M. M. (2022). A comparative study of soil liquefaction assessment using machine learning models. Geotechnical and Geological Engineering, 40(9), 4721-4734.