Optimizing computations in weather and climate prediction models

Abstract

To meet the challenge of developing a comprehensive weather and climate prediction model which can give realistic scenarios for many time scales, more computer power than is currently available will be needed. One possibility for alleviating this shortcoming is to increase the integration timestep. We propose and test several methods which may prove useful. One procedure is an expansion of the model dependent variables in a Taylor series. Application of this method to simple models indicates acceptable increases in timestep by a factor of five. A multi-level approach which is less complex to apply gives comparable results and is more successful when high accuracy is desired. To bypass the limiting constraint of the Courant-Friedrichs-Lewy (CFL) condition on gravity waves, an approach is suggested in which the prediction model is represented in its normal modes and the high frequency modes are balanced while the low frequency modes are predicted. Experiments with this procedure are described and in combination with the multi-level integration technique show substantial increases in integration timestep for acceptable integration results, both on the forecast and climate scale. Experiments are now underway applying this process to the NCAR/CCM3, a state-of-the-art model.