Reducing Uncertainty of Prediction from Empirical Correlations

Abstract

Empirical correlations with in situ or laboratory test results are routinely used in geotechnical designs and the uncertainty of prediction from these correlations can be substantial at times. This paper formalizes a procedure for characterizing and analyzing uncertainty in a design in which linear correlations have been used. In particular, how this design uncertainty can be reduced from gathering regional data and site-specific observations is addressed. The Bayesian approach is then applied to update empirical predictions with regional information or site-specific observations and hence to effectively reduce the uncertainty associated with the correlation. Three levels of uncertainty reduction associated with the use of an empirical correlation for design are discussed based on the availability of regional data and site-specific observations. Analyses of correlations for the shaft resistance of rock-socketed piles based on a global data set and a regional data set and for the shaft resistance of large-diameter bored piles in saporilitic soils based on site-specific observations are demonstrated as examples to implement the proposed procedure. In both examples, the uncertainty with the predictions updated by the proposed approach is significantly smaller than those directly associated with empirical correlations.