Uncertainty Quantification and Learning in Geophysical Modeling: How Information is Coded into Dynamical Models

Abstract:

There is a clear need for comprehensive quantification of simulation uncertainty when using geophysical models to support and inform decision-making. Further, it is clear that the nature of such uncertainty depends on the quality of information in (a) the forcing data (driver information), (b) the model code (prior information), and (c) the specific values of inferred model components that localize the model to the system of interest (inferred information). Of course, the relative quality of each varies with geophysical discipline and specific application.

In this talk I will discuss a structured approach to characterizing how ‘Information’, and hence ‘Uncertainty’, is coded into the structures of physics-based geophysical models. I propose that a better understanding of what is meant by “Information”, and how it is embodied in models and data, can offer a structured (less ad-hoc), robust and insightful basis for diagnostic learning through the model-data juxtaposition. In some fields, a natural consequence may be to emphasize the a priori role of System Architecture (Process Modeling) over that of the selection of System Parameterization, thereby emphasizing the more creative aspect of scientific investigation – the use of models for Discovery and Learning.