Data cleaning and feature selection for gravelly soil liquefaction

Abstract

Abstract Liquefaction of gravelly soil has been reported for several historical earthquakes. However, the data size remains insufficient for guaranteeing a high-performance prediction model, especially because the data quality used for the model building has not been evaluated in previous studies. In addition, the significant factors used to construct a gravelly soil liquefaction model remain unclear. To overcome these issues, the following key efforts are made in this study: (1) significantly expanded databases are accumulated for filed performance case histories obtained using dynamic penetration and shear wave velocity tests; (2) the data quality is improved by screening, correction, and repair of filed data case histories; (3) a framework is proposed to identify significant factors for gravelly soil liquefaction; and (4) the thresholds for two triggers of gravelly soil liquefaction are updated as Hn (the impermeable capping layer) larger than 0\xa0m and Dn (the thickness of the unsaturated zone between the groundwater table and the capping layer) less than or equal to 4\xa0m. Data cleaning and identification of significant factors can both improve the predictive performance of a model.