In the absence of comprehensive observations, initial conditions and forcing fields for models of the ocean and atmosphere must be estimated. In the case of numerical weather prediction, initial conditions for continuing forecasts are estimated by combining current observations with results of previous forecasts. This process of estimating the state of the system by combining model output with observation is known as “data assimilation.”
Ocean models intended for coupling to atmospheric models for climate prediction or coupling to chemical or biological models are typically coarsely resolved due to limitations on computing resources, and are not able to simulate observed phenomena faithfully. Major ocean features such as the Gulf Stream are often systematically misrepresented in position and strength, so attempts to assimilate observations of the Gulf Stream into coarsely resolved models result in systematic errors, as the model adjusts to the new input by internal model dynamics that have no counterparts in nature.
Variability in observed data due to physical phenomena that are not adequately modeled is known as “representation error,” and cannot be usefully assimilated. We propose a method for distinguishing observed variability that can be modeled from that which cannot, and present preliminary results to demonstrate its effectiveness.