Development of Bayesian Probabilistic Approaches for Rock Property Characterization

Abstract

The design and analysis of rock structures require parameters of rock properties, which makes the determination of rock properties an essential task in rock engineering. Characterization of rock properties relies on several methods like in-situ and/or laboratory tests and empirical estimations. Various uncertainties are unavoidably involved in rock property characterization, such as inherent variability, measurement errors, model uncertainty and statistical uncertainty. These uncertainties have not been explicitly considered in the traditional selection or use of empirical models and/or methods, and in deterministic rock property characterization. With probabilistic methods gaining popularity in rock mechanics and rock engineering, there is a need to rationally quantify these uncertainties through probabilistic characterization of rock properties, in order to provide input data for probability-based rock engineering design and analysis. This study aims to address these challenges through the development of Bayesian probabilistic methods for rock property characterization.
A Bayesian framework is presented for rock property characterization, which integrates systematically prior information and site-specific test results to select empirical models for specific sites, and probabilistically characterize rock properties and their correlation. The Bayesian framework explicitly addresses the inherent variability of rocks, and rationally accounts for other uncertainties arising from rock characterization. Next, a database is developed from a global compilation of rock properties reported in the literature for the three types of rock (i.e., igneous, sedimentary and metamorphic rock). The typical ranges of rock property data, range and mean of rock property mean, and range and mean of rock property coefficient of variation (COV) of relevant rock properties are evaluated and summarized for the three rock types. The typical ranges of rock can serve as approximations when rock data are not available at a site, and as prior information in the Bayesian framework during rock property characterization.
Bayesian approaches are developed to select the most appropriate model for specific sites, among sets of candidate models. The Bayesian approaches integrate site test data with prior knowledge to select the most appropriate model for characterization of one or more rock properties and their correlation. Next, the Bayesian framework is integrated with Markov chain Monte Carlo (MCMC) simulation to develop approaches for generating a large number of equivalent samples of the rock properties concerned for probabilistic characterization. When only sparse test values of a rock property (e.g., point load index, Is(50) ) are available from a site, the Bayesian approach integrates the limited
available data and prior information to generate a large number of equivalent samples of the rock property concerned (e.g., uniaxial compressive strength, UCS) for its probabilistic characterization. When sparse test values of more than one rock property (e.g., UCS and Young’s modulus (E) from compression tests) are available from a site, the Bayesian approach integrates the prior information with the limited test values of the properties, to generate large sample pairs of the rock properties concerned for probabilistic characterization of their correlation. The approaches are illustrated and validated with several sets of real observation and simulated data. It is shown that they provide proper probabilistic characterization of rock properties and their correlation.
Approaches are further developed based on the Bayesian framework to integrate site observation data, with quantification or guideline charts (e.g., geological strength index (GSI) quantification chart, Hoek-Brown constant guideline table) commonly used in estimating rock properties, to evaluate uncertainties in the charts and provide full probability distributions of the rock properties concerned. The approaches are illustrated with sets of real observation data, and are found to properly quantify the uncertainties.

Date of Award	8 Aug 2016
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Yu WANG (Supervisor)