Machine learning algorithms-based evaluation of dynamic properties and liquefaction potential of soil

Abstract

Machine learning models are the viable option contrary to the conventional experimental procedures that are complicated and tedious for liquefaction potential prediction. This study leverages the strength of machine learning models including K-nearest neighbor (KNN), random forest (RF), gradient boost (GB), extreme gradient boost (XGB), decision tree (DT) and artificial neural network (ANN) for predicting the soil responses and liquefaction susceptibility by utilizing the 500 SPT-N cases as input dataset. The predictive capabilities of models in assessing the intricate relationship between various soils parameters are assessed by employing the evaluation matrices of mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), Nash-Sutcliffe efficiency, Percent Bias, Weighted Index and R2. Moreover, the comparative analysis of the models has been conducted to find the best optimized model for liquefaction potential determination by accuracy matrix, Akaike information criterion (AIC) and ranking analysis. The results indicated that RF model exhibited the highest prediction efficiency for shear stress with R2 of 0.998. Similarly, GB indicated the better performance as compared to other models in evaluating the shear wave velocity with 0.992 coefficient of determination (R2). The most important and critical parameter for liquefaction maximum shear modulus (Gmax) is predicted with high accuracy by all models with outperformance of ANN having R2 value of 0.999. Moreover, liquefaction potential index (LPI) is also well predicted by XGBoost model. The performance trend of all the models for each of the parameters is also in accordance with AIC criteria. Furthermore, scaling analysis of all models for collective comparison put RF model on rank first for overall prediction of liquefaction potential and dynamic properties. XGBoost and DT showed second and third best performance, followed by GB as fourth, ANN as fifth and KNN as least effective model for this task. The findings of this study offer sophisticated models for evaluating soil behavior and liquefaction potential, which has important ramifications for geotechnical engineering.