Experimental Design : Optimizing Quantities of Interest to Reliably Reduce the Uncertainty in Model Input Parameters
As stakeholders and policy makers increasingly rely upon quantitative predictions from advanced computational models, a problem of fundamental importance is the quantification and reduction of uncertainties in both model inputs and output data. The typical work-flow in the end-to-end quantification of uncertainties requires first formulating and solving stochastic inverse problems (SIPs) using output data on available quantities of interest (QoI). The solution to a SIP is often written in terms of a probability measure, or density, on the space of model inputs. Then, we can formulate and solve a stochastic forward problem (SFP) where the uncertainty on model inputs is propagated through the model to make quantitative predictions on either unobservable or future QoI data. In this work, we use a measure-theoretic framework to formulate and solve both SIPs and SFPs. From this perspective, we quantify the geometric characteristics of using hypothetical sets of QoI that describe both the precision and accuracy in solutions to both the SIP and SFP. This leads to a natural definition of the optimal experimental design, i.e., what the optimal configuration a finite set of sensors is in space-time. Several numerical examples and applications are discussed including the ADvanced CIRCulation (ADCIRC) model using simulated data from Hurricane Gustav to determine an optimal placement of buoys in the Gulf of Mexico to capture high water marks of the storm surge in order to reduce uncertainties in bottom roughness characteristics.
- Pub Date:
- January 2016
- Mathematics - Numerical Analysis