Ensemble-based soil liquefaction assessment: Leveraging CPT data for enhanced predictions

Abstract

This study focuses on predicting soil liquefaction, a critical phenomenon that can significantly impact the stability and safety of structures during seismic events. Accurate liquefaction assessment is vital for geotechnical engineering, as it informs the design and mitigation strategies needed to safeguard infrastructure and reduce the risk of catastrophic failures. To enhance the accuracy of classification problems associated with liquefaction, we employ ensemble methods, leveraging diverse machine learning techniques such as support vector machines, stochastic gradient descent, multi-layer perceptron neural networks, K-nearest neighbors, and decision trees. The research encompasses data exploration and a subsequent division for performance assessment, followed by hyperparameter tuning through GridSearchCV to optimize model effectiveness. Among the ensemble methods employed, AdaBoost stands out as the most accurate, achieving precision of 85%, recall of 84%, F1 score of 83%, Jaccard index of 72%, and overall accuracy of 84%. However, K-nearest neighbors and decision trees exhibit higher false negative values compared to other methods. Notably, both ensemble approaches provide acceptable estimations, with false negative values ranging from 0 to 1 and false positive values between 7 and 10. The decision tree, while predicting the lowest false positive rate, has a higher false negative count, rendering it less favorable for practical applications.