A comparative analysis of ensemble learning algorithms with hyperparameter optimization for soil liquefaction prediction

Accurate prediction of soil liquefaction potential is crucial for evaluating the stability of structures in earthquake regions. This study focuses on predicting soil liquefaction using a dataset that included historical liquefaction cases from the 1999 Turkey and Taiwan earthquakes. The dataset was divided into three subsets: Dataset A (fine-grained), Dataset B (coarse-grained), and Dataset C (all samples). Through the analysis of these subsets, the study aims to assess the performance of machine learning algorithms in predicting soil liquefaction potential. This study applied ensemble machine learning algorithms, including extreme gradient boosting, adaptive boosting, extra trees, bagging classifiers, light gradient boosting machine, and random forest, to accurately classify the liquefaction potential of fine-grained and coarse-grained soils. A comparison between the genetic algorithm approach for hyperparameter optimization and traditional methods such as grid search and random search revealed that genetic algorithms outperformed both in terms of average test and train accuracy. Specifically, the light gradient boosting machine yielded the best predictions of soil liquefaction potential among the algorithms tested. The study demonstrated that Dataset B achieved the highest learning performance with accuracy of 0.92 on both the test and training sets. Furthermore, Dataset A showed a training accuracy of 0.88 and a test accuracy of 0.84, while Dataset C exhibited a training accuracy of 0.87 and a test accuracy of 0.87. Future studies could build on these findings by evaluating the performance of genetic algorithms on a wider range of machine learning algorithms and datasets, thus advancing our understanding of soil liquefaction prediction and its implications for geotechnical engineering.