1.1
Development of neural network convection parameterizations for climate models using CRM simulations and ARM data
Vladimir M. Krasnopolsky, IMSG at NCEP/NWS/NOAA, Camp Springs, MD; and M. S. Fox-Rabinovitz, P. Rasch, Y. Kogan, and A. Belochitski
The neural network (NN) technique is used for development of novel fast convection parameterizations for climate models based on data from CRM (Cloud Resolving Model) simulations initialized with DOE ARM data. NN serves as an interface transferring sub-grid scale information from fine scale models (CRMs) into larger scale GCMs (upscaling). The System for Atmospheric Modeling (SAM) provided by M. Khairoutdinov (Khairoutdinov and Randall, 2003), has been used for CRM simulations. The NN methodological approach has been previously applied by investigators for NN emulations of climate model radiation (Krasnopolsky et al. 2008). ARM data is used for: (a) initializing SAM/CRM simulations, (b) developing NN convection parameterizations, and (c) validating model simulations.
The research is based on two major assumptions: (1) that SAM/CRM (initialized and/or driven by ARM forcing) simulated data and observed ARM data provide a more accurate representation of clouds and convection processes than existing convection parameterizations in GCMs, and (2) that NNs can extract and emulate this more accurate physics through learning from SAM/CRM simulated data and ARM data with the accuracy sufficient for the subsequent use of the trained NN as a convection parameterization in the NCAR SCM (Single Column Model) and/or GCM.
We are pursuing the following avenues of research. SAM/CRM simulated data have been produced with perturbed initial conditions for ensemble simulations and is used for creating data sets for NN training. Specifically, the SAM/CRM simulated data using the ARM or TOGA-COARE forcing for the 256 x 256 km domain with 1 km resolution and 96 vertical layers (0 – 28 km) is spatially averaged at every hour of model integration to produce NN training data sets with different effective resolutions of 256, 128, 64, 32, 16, or 8 km. It allows us to produce the number of cases/events needed for obtaining representative data sets for NN training.
The choice of optimal spatial averaging most suitable for NN training is being investigated. Then we train the NN convection parameterizations using the optimal spatially averaged data sets for the same inputs and outputs as those used in NCAR SCM (Single Column Model) CAM or NCAR CAM. The accuracy of developed NN convection parameterizations is validated using ARM or TOGA-COARE data. At the next step of the project, these NN convection parameterizations will be included into the NCAR SCM CAM, and tested in climate simulations using data from the SGP (Southern Great Plains) and TWP (Tropical West Pacific) ARM sites and TOGA-COARE data.
Acknowledgments: The authors would like to thank Prof. Marat Khairoutdinov (SUNY) for providing SAM and consultations, and Dr. Peter Blossey (UWA) for consultations on SAM.
References:
Khairoutdinov, M. F., and D. A. Randall, 2003: Cloud resolving modeling of the ARM summer 1997 IOP: Model formulation, results, uncertainties, and sensitivities. J. Atmos. Sci., 60, 607–625.
V. M. Krasnopolsky, M.S. Fox-Rabinovitz, and A. A. Belochitski, 2008: "Decadal Climate Simulations Using Accurate and Fast Neural Network Emulation of Full, Long- and Short Wave, Radiation.", Monthly Weather Review, 136,
Session 1, Applications of Artificial Intelligence Methods to Problems in Environmental Science: Part I
Tuesday, 19 January 2010, 1:30 PM-3:00 PM, B204
Next paper