Preliminary Liquefaction Susceptibility Using Different Machine Learning Techniques

Abstract

Liquefaction Analysis is one of the most important parameters in the study of geotechnical earthquake engineering. Till the 1960’s this phenomenon was relatively less unearthed, and practitioners have not formulated any detailed method for its assessment. However, two large earthquakes in 1964 in Niigata, Japan, and Alaska, USA have turned the attention of engineers toward this issue. Several laboratory and field methods were developed for the calculation of the Factor of Safety. Developed methods have been used all over the world which use the Cyclic Resistance Ratio and Cyclic Stress Ratio for measuring the factor of safety. As empirical meth-ods have inherent limitations due to certain assumptions for the ease of work, liquefaction for-mulations have also been generalized to counter the real-world scenario. These assumptions significantly impacted the implementation of the simplified methods. In recent years Machine Learning has been used to improve the inherent shortcomings present in the conventional method of predicting liquefaction. As machine learning algorithms learn from the data and are not explicitly programmed, they can develop highly non-linear relationships and learn from the data. Researchers have mostly used the published data based on the Factor of Safety to predict the liquefaction potential which is not a credible approach as it has inherent shortcomings in ascertaining CSR and CRR(Kurnaz & Kaya, 2019a).In this research work four different ma-chine learning models namely Logistic regression, Support vector machine, Decision tree, and Artificial neural networks have been used to predict the liquefaction potential of the soil based on the published field data. The data has been procured from credible published research papers. It includes six different input parameters named cone tip resistance, sleeve friction ratio, effec-tive stress, total stress, maximum horizontal ground surface acceleration, earthquake moment magnitude, and one output parameter named liquefaction. The performance of the developed models was gauged with the help of classification assessment report parameters named accu-racy, precision, recall, and F1 score. It was found that the Decision Tree algorithm performance x was the best among all the other algorithms followed by Artificial neural networks, Logistic regression, and Support vector machine. The developed models can be used as a predictive model for the preliminary liquefaction assessment of soil.