 |
Weather forecasting, modeling, and research are advancing rapidly with the advent of high performance computing and communications systems. Research groups at several agencies are moving traditional models to scalable systems. These models are then used to test the validity of the current understanding of the physics of weather and to develop more detailed, robust models. When the models are sufficiently trustworthy, they are used for operational forecasting by the National Weather Service and to drive air quality assessments for environmental analysts. |
Perhaps the most dramatic example of recent advances is the hurricane modeling system that became operational with the 1995 hurricane season as the culmination of a long development effort. The operational implementation of this model was made possible only by the use of the fastest computers available to the National Weather Service. When moved to a parallel operational setting, a speed-up of 18 over a serial implementation of the model was realized. Without this speed-up, forecasts could not be made soon enough to provide timely information. This is a clear example of the advantages of modern high performance computing applied to a problem which affects each of us every day. Further improvements are expected from better initialization procedures, more flexible grid design, and, of course, faster computers.
 |
The average forecast track error for the 1995 Atlantic hurricane season from the new forcast system developed by GFDL (red line) contrasted with the range of forecast track errors from earlier models. |
Our daily weather forecasts start as initial value problems on the National Weather Service supercomputer. Satellite, radar and balloon observations are assimilated and provide initial conditions for computer models of the atmosphere. Since these models have approximately a million degrees of freedom, and must be integrated with short time steps (minutes) for two weeks, it is essential to design the models in such a way that they can be efficiently implemented on Massively Parallel Computers (MPP). Since the current generation of models was designed for sequential vector architectures, transferring them to MPP or scalable parallel architectures is a challenging task.
Scientists at several National Laboratories and Supercomputer centers, and the center for Analysis and Prediction of Storms are investigating the application of scalable systems to the problems of mesoscale weather forecasting. These systems are exploring phenomena such as tornados, thunderstorms, and flash floods.
The National centers for Environmental Prediction are examining two approaches to global weather prediction: ensemble forecasting for determining the robustness of forecasts, and a "conformal expanded cube" that may be more efficient to parallelize than the traditional spectral models.
 |
Sample output from an MPP implementation of a weather model showing the improved level of detail obtainable through one of these models. The figure may show state-level detail not available in a traditional model. |
 |
 |