Application of two-level factorial design to sensitivity analysis of keyblock statistics from fracture geometry

Abstract

A stochastically modeled fracture network offers potential for more realistic assessment of stability status in underground excavations than predictions based entirely on deterministic features. The reliability of probabilistic models, however, depends strongly on an accurate estimation of the model's variables, i.e., the fracture network properties from the field and laboratory observations. In this study, predictions of keyblocks by implementing stochastically generated fractures in the Central Storage Facility for Spent Nuclear Fuel (CLAB 2 Centralt Lager Använt Bränsle) located in southeast Sweden are presented. The fracture network model is built by using fracture mapping in the floor of the facility and incorporates fracture size, shape, orientations, termination mode, spatial arrangement and fracture mechanical properties. The predicted volume of individual keyblocks is best-fitted with the Pareto probability distribution function. Subsequently, a statistical two-level factorial analysis is performed to examine the impact of both single fracture properties and their interactions on the predictions made. In the factorial experiments, the block predictions are made for eight different fracture models where three factors: fracture radii, orientation and termination are each assigned two levels intentionally departing from the best estimates found for the CLAB 2 site. This allows us to express the experiment results as the degree to which each of the eight computed block statistics deviated from the most likely prediction. It is found that fracture orientation is the only statistically significant factor influencing the keyblock statistics while the input from other variables/fracture properties and their interactions is less significant. The results of our study yield a prospective approach for improving the effectiveness of the stochastic model variable estimation and for more optimal field mapping strategies.