Novel methodologies for soil characterization from CPT data

Abstract

This paper presents results and comparisons of soil characterization from cone penetration test (CPT) data using some of the traditional and nontraditional soil classification methods. Recently, a general regression neural network (GRNN) model was developed for predicting soil composition (percent sand-, silt-, and claysized particles) from CPT data. CPT data, together with grain size distribution results of soil samples retrieved from adjacent SPT boreholes from several sites in Taiwan following the Chi-Chi earthquake, were used to train and test the network. The trained GRNN was validated with previously unseen data and the model predictions were compared with the reference particle size distributions and the soil behavior type CPT soil classification method proposed by Robertson. The results were also compared to a nontraditional statistical soil classification approach, termed probabilistic region estimation that was proposed by Zhang and Tumay to estimate the probability of sand, silt, and clay in soils.