Steven Cohn (PI)
NASA Goddard Space Flight Center
Steven.Cohn@nasa.gov

A Data Assimilation Approach Based on Physical First Principles

Successful utilization of the enormous volume and variety of data now becoming available from new satellite sensors presents major, pressing scientific and technological challenges to the NASA community. Among these challenges is the development of data assimilation technology sufficiently powerful to reap quickly the full potential impact of the observations. The main recent thrust in this direction has been the development of fully four-dimensional data assimilation schemes based on discrete Kalman filtering. Despite theoretical expectations, major implementation efforts at leading national and international data assimilation centers have so far failed to yield improvement upon the impact of observations obtained from benchmark operational 3D schemes, and at least one well-known effort has been abandoned after years of work. The reason for this failure has been shown, finally, to arise from a major flaw in the traditional discrete least-squares estimation framework itself, on which all of these schemes are based, rather than on details of any particular 4D scheme. Three-dimensional schemes are also based on this framework, and the same flaw has been linked to known but hitherto unexplained deficiencies in the impact of observations that is obtained with 3D schemes.

The flaw is that the traditional framework is fundamentally inconsistent with the basic integral conservation laws of continuum physics: conservation of mass, energy and momentum. It has been corrected by posing data assimilation directly as a problem in continuum physics, rather than as a fundamentally discrete and probabilistic one. The perspective of the corrected framework is conceptually very close to that of the main direction in engineering disciplines, which have made rapid progress on critical estimation and control problems over the past decade or so by discarding the traditional (but generally unverifiable) probabilistic assumptions and replacing them by a deterministic optimality criterion. The corrected, physically consistent framework gives rise to an entirely new approach to data assimilation. This approach goes directly to the heart of what some researchers have long suggested is the main practical difficulty with 4D data assimilation methods; that it is not their excessive computational requirements per se, but rather that they require specification of and excessive amount of statistical information which is not available.

We propose an effort to carry out an experimental development, implementation, testing and validation of this physically-based approach. We expect the proposed effort to place GMAO in a position to propose a major development effort along these lines for its core data assimilation system five years from now.

+ Back to Participant Listing