Abstract:
|
The conventional construction of Gaussian process (GP) emulators assumes a scalar output that is a stationary function of a modest number of input parameters. It is of great interest to construct emulators for high dimensional input and output spaces. A different but related topic of interest is uncertainty quantification of coupled models – that is, when a functional of the output from one model serves as part of the inputs for a second model. In particular this kind of coupling arises in multi-scale and multi-physics computer models, e.g. when specialized computer codes are used to simulate different parts of the physics of a highly complex mathematical model. We will describe the construction of linked GP emulators of fluid flow coupled to the mechanics of a deformable porous material where specialized codes operate on the fluid and solids portions independently, and are coupled by an exchange of high-dimensional output data from each individual model.
|