Dynamics characterization and prediction of seismic-induced soil liquefaction

Abstract

Seismic-induced soil liquefaction takes place when the propagation of seismic waves weakens the soil's ability to hold shear stress and increases pore water pressure, causing a change in rheology and leading to liquefaction. By employing supervised machine learning (ML), a liquefaction potential assessment was developed using a standard penetration test (SPT) dataset. CatBoost is identified as the best-performing model among 13 supervised ML classifiers in 15-fold stratified cross-validation, with an average accuracy of 85.47%. Via ExtraTrees, which was trained under the entropy criterion using the SPT dataset, and Pearson correlation coefficients, the most significant contributing factors to seismic-induced soil liquefaction are standard penetration resistance, normalized cyclic stress ratio, shear wave velocity, magnitude scaling factor, and unit weight below the groundwater table. In 20 train-test runs, CatBoost (in its default setting) along with the five features yielded 85.89% accuracy. The f1, recall, and precision scores are 86.90%, 85.79%, and 88.52%, respectively. This study also presents a "six-dimensional" visualization that aims to characterize the seismic-induced soil liquefaction phenomenon.