A Stochastic Inverse Problem for Multiscale Models

A Stochastic Inverse Problem for Multiscale Models

Abstract

Descriptions of complex multiscaled physical systems often involve many physical processes interacting through a multitude of scales. In many cases, the primary interest lies in predicting behavior of the system at the macroscale (i.e., engineering scale) where continuum, physics-based, models such as partial differential equations provide high-fidelity descriptions. However, in multiscale systems, the behavior of continuum models can depend strongly on microscale properties and effects, which are often included in the macroscale model as a parameter field obtained by some upscaling process from a microscale model. Generally, a number of choices have to be made in choosing an upscaling procedure and the resulting representation of the parameter. These choices have a strong impact on both the fidelity and the computational efficiency of the model. Thus, choosing a good parameter representation and upscaling procedure becomes part of the uncertainty quantification and prediction problem for a multiscale model. We consider the use of output data from the macroscale model to formulate and solve a stochastic inverse problem to determine probability information about the upscaled parameter field. In particular, we extend a measure-theoretic inverse problem frame-work and non-intrusive sample-based algorithm to determine the choices of parameter representation and upscaling procedure that are most probable given uncertain data from the macroscale model.We illustrate the methodology in the context of shallow water flow and sub-surface flow.

Publication
In International Journal for Multiscale Computational Engineering