Effect of Model Uncertainty on Probabilistic Characterization of Soil Property

摘要

Reliability-based design and analysis in geotechnical engineering requires input parameters, such as soil properties, to be probabilistically characterized. This generally requires a large number of site-specific data. However, site-specific data is often sparse and limited, particularly for geotechnical projects with small to medium sizes. To facilitate the probabilistic characterization of soil property of interest (e.g., effective friction angle, of soil), Bayesian equivalent sample approach has been developed. It systematically integrates limited site-specific data with engineering judgment/local experience (i.e., prior knowledge in Bayesian methods) and regression models (relating soil properties to site-specific data, if the soil properties of interest are not measured directly). As the regression model (e.g., a commonly used design chart between standard penetration test (SPT) data N_(SPT) and ф') is generally not perfect but with some uncertainty, the characterization result would be inevitably affected by the uncertainty in the regression model. Furthermore, the effect of model uncertainty may become more sophisticated, if the magnitude of model uncertainty in regression models (e.g., a N_(SPT) - ф' design chart) is unknown or difficult to calibrate. This paper aims to explore the effect of model uncertainty on the characterization result, particularly when the magnitude of model uncertainty is unknown (note that determination and quantification of the model uncertainty are not the objective of this study). The effect of model uncertainty can be clearly illustrated by comparing the probabilistic characterization result of ф' considering the unknown model uncertainty in a N_(SPT) - ф' design chart, and that ignoring the unknown model uncertainty in the N_(SPT) - ф' design chart. Simulated data is used for such illustration. It is shown that considering the model uncertainty in the design chart achieves more consistent and reliable results than ignoring model uncertainty in the design chart. This would be quite useful when probabilistically estimating soil properties of interest (e.g., ф') from some other commonly used in-situ tests (e.g., N_(SPT)).