Approaches to highly parameterized inversion: A guide to using PEST for model-parameter and predictive-uncertainty analysis

Abstract

mined problems, both linear and nonlinear. PEST tools for calculating contributions to model predictive uncertainty, as well as optimization of data acquisition for reducing parameter and predictive uncertainty, are presented. The appendixes list the relevant PEST variables, files, and utilities required for the analyses described in the document.

Approaches to highly parameterized inversion-A guide to using PEST for groundwater-model calibration

“…It also indicates that the degree of nonuniqueness would be greater if the possible parameter fields used to make the prediction included additional geologically realistic complexity not supported by the current calibration data. Such an encompassing estimate of predictive uncertainty can be calculated by using the null space Monte Carlo method (Tonkin and Doherty, 2009;Doherty, Hunt, and Tonkin, 2010). In this analysis, the uncertainty associated with any prediction that depends on realistic parameter detail that cannot be discerned from the calibration data will be higher than that calculated by selection of any one simplified parameter field.…”

“…Utilities available through the PEST suite include several postprocessing functions that can add value to the outcomes of the parameter-estimation process (for example, see Doherty, Hunt, and Tonkin, 2010). In fact, some of these tasks can be done before calibration is begun by using only precalibration parameter sensitivities calculated, for example, from a PEST run with NOPTMAX set to −1 or −2.…”

“…Finally, several postcalibration utilities can be used to formally investigate model parameter predictive uncertainty, as well as the worth of future data collection and network design for reducing this uncertainty. Approaches for applying these methods to highly parameterized problems are discussed in detail in Doherty, Hunt, and Tonkin (2010). The following tables include a column labelled "usage," which contains an index whose value is between 1 and 3.…”

Enterprise engineering and management at the crossroads

“…This is a typical situation in environmental modelling and management: the most successful environmental models used for decision making are limiting themselves to local management issues on smaller system scales, and global environmental modelling is still grappling with the problem of validation [58].…”

RRAWFLOW: Rainfall-Response Aquifer and Watershed Flow Model (v1.15)

“…Although Vapnik-Chervonenkis generalization theory suggests that models with higher complexity tend to have higher prediction uncertainty, model complexity is not necessarily proportional to prediction uncertainty (Fienen et al, 2010). Doherty et al (2010) described a method for predictive uncertainty and sensitivity that tests the range of each parameter's potential values on the basis of expert knowledge and propagates this uncertainty to model predictions. Estimating this potential range of IRF parameter values, however, might be more difficult than, for example, estimating hydraulic conductivity or streambed roughness in a distributed model.…”