NGWA Event and Education Recordings Archive

Sensitivity and uncertainty analysis provide insights into and reveal consequences of the often complex set of processes needed to simulate groundwater and other environmental systems. Sensitivity analysis identifies observations important to parameters, parameters important to predictions, and observations important to predictions. Uncertainty analysis quantifies the precision with which predictions are calculated given the observations and other knowledge available for model development, including model construction. Model construction uncertainties can be evaluated using multiple alternative models that use alternative boundary conditions, numerical formulations, computer codes, and so on. Alternative models and more complicated models can be explored when sensitivity and uncertainty analyses are conducted using computationally frugal methods that require fewer model runs amenable to high performance computing. In this talk, recent investigations of computationally frugal methods are reviewed, including comparison with the computationally demanding Sobol and cross-validation methods. The new hybrid DELSA (Distributed Evaluation of Local Sensitivity Analysis) method is also discussed. Examples use groundwater and surface water models (MODFLOW, MODPATH-OBS, FUSE, and TOPKAPI). Results illustrate the insights that can be obtained from sensitivity and uncertainty analyses in general, and what is compromised and achieved using computationally frugal and demanding methods. Some examples include: (1) Differences between alternate groundwater models were more important than differences between local and Sobol measures of prediction uncertainty, suggesting that computational effort be focused on exploring alternative models. (2) A variance-based local method is used in a new way to display the information provided by observations for parameters from defined observation types. The local methods were found to be useful except for difficulties with FUSE models revealed using DELSA. Recommended methods and diagnostics for choosing between methods are discussed.