UMD AOSC Seminar
Using Neural Networks to Improve Atmospheric Model Physics
Dr. Vladimir Krasnopolsky
NOAA NCEP Environmental Modeling Center
University of Maryland, Earth System Science Interdisciplinary Center
Two related applications of neural networks (NN) to atmospheric model physics developed in collaboration with Dr. M. Fox-Rabinovitz and A. Belochitski (ESSIC) will be discussed:
1. An approach to accurate and fast calculation of model physics components using neural network emulations (NN) was developed and applied to full model radiation, the most time-consuming part of model physics. It has been implemented into the coupled high resolution NCEP Climate Forecast System. The developed highly accurate neural network emulations of long-wave and short-wave radiation parameterizations are 12 and 45 times faster than the original long-wave and short-wave radiation parameterizations, respectively. Comparisons of parallel decadal climate simulations and seasonal predictions performed with the original NCEP model radiation parameterizations and with their neural network emulations are presented. Almost identical results are obtained for the parallel decadal simulations.
2. The NN approach has been formulated for developing new parameterizations of model physics using data simulated by high resolution process models. It was applied to developing a new convection parameterization for NCAR CAM based on the data simulated by a Cloud Resolving Model. A prototype of the stochastic NN ensemble convection parameterization has been developed and initially tested in NCAR CAM. The obtained promising results are discussed and the future plans are outlined.
April 28, 2011, Thursday
AOSC 818. Frontiers in Atmosphere, Ocean, Climate, and Synoptic Meteorology Research