I am a graduate student in meteorology at MIT working with Professor Kerry Emanuel. I am working in the field of Numerical Weather Prediction, specifically Data Assimilation.

Abstract for NERSC ACTS Toolkit Workshop

It is commonly thought the next big break-through in numerical weather prediction (NWP) will come not from improved forecast models but from improved data assimilation systems. Forecast errors are unavoidable in NWP because of imperfections inherent in our observations of the atmosphere and in the models we use to predict the atmosphere's evolution. In the language of signal processing, data assimilation is a filtering problem where the truth state is periodically estimated from imperfect knowledge of the system. Typically this entails adjusting a background field with new information from observations of the state constrained by knowledge of error statistics whenever there are misfits between the background field and the new observations. The analyzed truth state then becomes the initial conditions for the next batch of forecasts.

Ensembles have been recognized as potentially useful to NWP since the late 1960s. Initially they were conceived of to better quantify the inherent errors in weather forecasts by making explicit use of the high sensitivity forecasts have on their initial conditions. An ensemble of states which differ by errors representative of the uncertainty believed to be in our analysis of the atmosphere is evolved in time and the divergence of its members clues us in to areas of the forecasted state where we have less confidence. Recently, ensembles have been shown useful in data assimilation methods as well. They help both to identitify and explore various dynamically interesting directions in which to perturb control forecasts (i.e. the directions is state space most likely to give sizable errors) and to estimate flow-dependent error statistics along the way. In the former, error growth limits as told by the divergence of ensemble members perturbed in the dynamically interesting directions (e.g. in the direction of approximate Lyapunov vectors or singular vectors) give information on where observational information should weighted more heavily (versus the initial background guess). In the latter, ensembles can be used to estimate the evolution of error-statistics without having to go through the computationally unfeasible process of explicitly evolving them via a Kalman filter or similar method.

We are researching novel data assimilation methods. A very important part of this is the exploration of new uses for and the refinement of the aforementioned uses for ensembles in data assimilation. We have begun by testing changes caused by assimilating different variables like potential vorticity and vortex locations. We plan to make further explorations into using different coordinates and different models like contour dynamics models. Hopefully our studies will provide us with information on the best uses for ensembles in data assimilation, with the anticipated spin-offs of how best to initialize ensembles and what size ensembles are necessary to do data assimilation well.

Numerical weather prediction has always been one of the premier computationally intensive fields. Over the years, forecast models have increased in complexity and resolution, limited mainly by the available computational resources. As data assimilation systems increase in complexity and as the need for larger ensemble sizes increases, the required computer resources will continue to increase. Currently, we are running simplified models to keep the cause and effect relationships in our experiments somewhat simple. Running experiments on a DEC alpha seems to be sufficient for the rough resolution and moderate ensemble sizes currently in use. However, as new data assimilation methods seem to prove useful in these simplified model contexts we will surely want to test them further in more complex (and realistic) models which will require greater power. Other groups around MIT have begun to build parallel clusters of PCs and these seem a very logical way to proceed for us, especially given the obvious parallel nature of running ensembles.