Probabilistic Analysis of the Compressibility of Soils

Abstract

Geotechnical engineers are always faced with uncertainties and spatial variations in material parameters. In this work, we propose to develop a framework able to account for different types of uncertainties in a formal and logical manner, to incorporate all available sources of information, and to integrate the uncertainty in an estimate of the probability. In geotechnical engineering, current soil classification charts based on CPT data may not provide an accurate prediction of soil type, even though soil classification is an essential component in the design process. As a cheaper and faster alternative to sample retrieval and testing, field methods such as the cone penetration test (CPT) can be used. A probabilistic soil classification approach is proposed here to improve soil classification based on CPT. The proposed approach provides a simple and straightforward tool that allows updating the soil classification charts based on sitespecific data. In general, settlements can be the result of surface loads or variable soil deposits. In current practice, the analysis to determine settlements is deterministic. It assumes that the soil profile at a site is uniform from location to location, and only allows limited consideration of the variations of the material properties and initial conditions within soil layers in spite of the wide range of compositions, gradations, and water contents in natural soils. A Bayesian methodology is used to develop an unbiased probabilistic model that accurately predicts the settlements and accounts for all the prevailing uncertainties. The proposed probabilistic model is used to estimate the settlements of the foundation of a structure in the Venice Lagoon, Italy. The conditional probability (fragility) of exceeding a specified settlement threshold for a given vertical pressure is estimated. A predictive fragility and confidence intervals are developed with special attention given to the treatment and quantification of aleatory and epistemic uncertainties. Sensitivity and importance measures are computed to identify the key parameters and random variables in the model.