Stochastic, goal-oriented rapid impact modeling of uncertainty and environmental impacts in poorly-sampled sites using ex-situ priors. (January 2018)

• Record Type: Journal Article
• Authors: Li, Xiaojun; Li, Yandong; Chang, Ching-Fu; Tan, Benjamin; Chen, Ziyang; Sege, Jon; Wang, Changhong; Rubin, Yoram
• Is Part Of: Advances in water resources. Volume 111(2018)
• Page Start: 174
• Page End: 191
• Publication Date: 2018-01
• Subjects: Hydrogeology -- Uncertainty -- Tunneling -- Environmental impacts -- Bayesian priors -- Site-characterization; Hydrology -- Periodicals; Hydrodynamics -- Periodicals; Hydraulic engineering -- Periodicals; 551.48
• DOI: 10.1016/j.advwatres.2017.11.008
• Languages: English
• ISSNs: 0309-1708

Abstract

Highlights: Stochastic modeling intends to accurately model uncertainty, not to reduce uncertainty. Stochastic analysis does not mandate minimum data requirements. A systematic approach for constructing priors from ex-situ data is presented. Utilization of scaling relationships allows modelers to avoid costly parameterization schemes. Costly estimates of parameters can be bypassed by focusing on impacts instead, and using the total probability theorem, and without loss of information. Abstract: Modeling of uncertainty associated with subsurface dynamics has long been a major research topic. Its significance is widely recognized for real-life applications. Despite the huge effort invested in the area, major obstacles still remain on the way from theory and applications. Particularly problematic here is the confusion between modeling uncertainty and modeling spatial variability, which translates into a (mis)conception, in fact an inconsistency, in that it suggests that modeling of uncertainty and modeling of spatial variability are equivalent, and as such, requiring a lot of data. This paper investigates this challenge against the backdrop of a 7 km, deep underground tunnel in China, where environmental impacts are of major concern. We approach the data challenge by pursuing a new concept for Rapid Impact Modeling (RIM), which bypasses altogether the need to estimate posterior distributions of model parameters, focusing instead on detailed stochastic modeling of impacts, conditional to all information available, including prior, ex-situ information and in-situ measurements as well. A foundational element of RIM is the construction of informative priors for target parameters using ex-situ data, relying on ensembles of well-documented sites, pre-screened for geological and hydrological similarity to the target site. The ensembles are built around two sets of similarity criteria: a physically-based set of criteria and an additional set covering epistemic criteria. In another variation to common Bayesian practice, we update the priors to obtain conditional distributions of the target (environmental impact) dependent variables and not the hydrological variables. This recognizes that goal-oriented site characterization is in many cases more useful in applications compared to parameter-oriented characterization.