P3.3
Model Parameter Evaluation using Linear Inverse Modeling

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Thursday, 2 February 2006
Model Parameter Evaluation using Linear Inverse Modeling
Exhibit Hall A2 (Georgia World Congress Center)
John A. Dykema, Harvard Univ., Cambridge, MA; and B. F. Farrell, S. S. Leroy, and J. G. Anderson

The key requirement for developing an accurate climate forecast is that the forecast model must contain the correct relationship between anticipated forcings and the response in climatological variables of interest. This relationship is not captured by the mean statistics of the climate system, but rather by the second moment statistics, as defined by the covariances of observable properties of the climate system. This concept derives from a formulation of the fluctuation-dissipation theorem applied to the climate by Leith (1975). We seek a methodology that extracts the required covariances from climate observations in a form which allows for the testing and improvement of climate models. This method is provided by Linear Inverse Modeling (LIM), a technique which extracts the dynamical relationship between climate variables directly from observations (Penland and Magorian 1993). The power of LIM derives from its ability to obtain dynamics from any well-observed, representative measurement of the climate system. It has been demonstrated in other climate applications using a variety of data sets. The LIM technique has been successfully employed to make forecasts of El Niņo events (Moore and Kleeman 1998). It also reproduces accurate statistics for climatologies of mid-latitude storm tracks (DelSole and Farrell 1996, Zhang and Held 1999, Whitaker and Sardeshmukh 1998). These results imply that LIM is fundamentally sound in capturing the dynamical relationships which determine the evolution of the climate system over decadal timescales. The correctly determined dynamical relationship among these variables provides a powerful constraint on the representation of the boundary layer in a climate model. This constraint must be satisfied for any model to forecast future climate credibly. Continued progress in model refinement requires developing methods to systematize parameterization improvement. The success of linear inverse modeling (LIM) for ENSO and seasonal climate forecasting (Penland and Magorian 1993; Winkler et al. 2001) reveals that inverting selected observations for dynamics is a powerful methodology compared with approximating dynamics of complex processes from first principles. This is particularly true when, as in the case of clouds and precipitation, the first principles are not themselves well known. One may have derived parameterizations for use in a model from first principles augmented by engineering formulae but it is very difficult to imagine being more precise in obtaining a model dynamics by such ab initio methods than by reconstructing the observed dynamics itself. A benchmark for model improvement, therefore, is that the model reproduce the LIM dynamics in appropriate variables. But more can be done than to systematize comparison among models and observations; a method can be developed to use such comparisons to directly improve models. What is needed is a way to evaluate model parameterizations, using observations, that tests whether the new parameterization improves the model's ability to forecast climate response to external forcing.