Max Suarez (PI)
NASA Goddard Space Flight Center
suarez@janus.gsfc.nasa.gov

Use of Neural Network Techniques in the Physics Parameterizations of Atmospheric General Circulation Models

As faster methods have been developed to compute the resolved dynamics in AGCMs (e.g., FVCORE), the cost of the physical parameterizations has become proportionately larger. Recent work (Krasnopolsky and Chevallier, 2003, Neural Networks, 16, 335-348 ) has shown that the cost of some parameterizations can be greatly reduced by using Neural Networks (NN)s) to approximate their results. Our submission will further develop the use of neural networks to approximate the results of the physically based parameterizations and apply them to the NASA GMAO model (GEOS-5). The research being proposed will be novel in several ways:

• It will use a Kalman-filter technique to train the network. This technique, although in wide-spread use in neural network algotithms, has only recently been tested in environmental problems (Lary and Mussa 2004, ACP-LOGO). Its main attraction is that it results in dramatically faster training of feed-forward networks than the usual back-propagation algorithm. 
  
• It will extend the application of NN methods to most of the model?s parameterizations. Most focus up to now has been on the radiation parameterizations, and the approach has been to simply take the outputs and inputs of the parameterization in GCM simulations and use them for training.  Extension to other parameterizations, such as boundary layer or moist processes will require a more nuanced approach where sub-problems within the parameterization are tackeld with the NNs while some of the physically based content is retained. 
  
• It will develop techniques for dynamic learning and quality control of the NN approximations. The application of NNs to models used for climate prediction naturally begs the question of whether the approximations obtained with one climate are applicable to a climate prediction. Our approach will be to develop methods to be continually training or retraining the parameterization in the course of the simulations and quality controlling its results by producing targeted sampling results with the full parameterization. 
  
• Explore the use of NNs to directly construct empirical parameterizations from a combination of observations and off-line cloud-resolving simulations.   
  
• Develop the NN parameterizations with a view toward their use in data assimilation and other  applications that may require an adjoint of the atmospheric model. 

+ Back to Participant Listing