A new approach based on application of advanced neural network (NN) techniques to accurate and fast emulation of model physics components for NCAR CAM, has been introduced and preliminary investigated by the authors for acceleration of calculating physical processes in atmospheric (Krasnopolsky et al. 2004a-d, Fox-Rabinovitz et al. 2004) and ocean (Krasnopolsky et al. 2002) models. This approach is a novel synergetic combination of NN techniques and deterministic modeling. NN emulations have been developed for an existing parameterization of atmospheric physics that has already been very carefully tested and validated by its developers off-line and then on-line through experimentation with the entire model. More specifically, NN emulations for the NCAR CAM long-wave radiation (LWR) (Collins 2001, 2002) have been developed and tested. A similar development for the NCAR CAM short-wave radiation is under way. Due to the capability of NNs to provide an unprecedented accuracy for approximation of complex multidimensional systems like model physics, our NN emulations of model physics parameterizations are practically identical to original physical parameterizations with the accuracy and speed-up of NN emulations always measured against the original parameterization. This allows us to preserve the integrity and the level of sophistication of the state-of-the-art parameterizations of physical atmospheric processes. The developed NN emulation for LWR is 80 times faster than the original LWR parameterization. For assessing the impact of using NN emulation of LWR for climate modeling, the parallel NCAR CAM decadal climate simulations have been performed with the original LWR parameterization (the control run) and with its NN emulations. Comparisons of the control and NN emulation runs are done by analyzing the decadal mean differences or biases between the parallel run prognostics and diagnostics. In the climate simulations performed with the original LWR parameterization and its NN emulations the time mean surface pressure is virtually identical. Namely, there is a negligible difference of 0.0001% between the NN run vs. the control run. Other time global means for model diagnostics and prognostics are also very close and show a profound similarity between the simulations, with the differences within about 0.03 – 0.1% and not exceeding 0.3%. Most importantly, biases for decadal simulations are not accumulating during the model integration. The LWR cooling rates obtained in these climate simulations maintain the accuracy of approximation (biases and RMSE) consistent with the accuracy obtained for the NN training period with very small systematic errors or biases and small random errors, well within observation errors. Even the maximum biases for zonal and time mean prognostic fields such as temperature, wind and specific humidity for the NN run are 0.1-0.2 K, 0.1-0.2 m/s, 0.01-0.02 g/kg - only a small fraction of observation errors.

The horizontal distributions of time means as well as time series of global and regional means for prognostic and diagnostic fields are also close to each other, with the maximum differences within a half of observation errors.

REFERENCES: Collins, W.D., 2001: Parameterization of generalized cloud overlap for radiative calculations in general circulation models, J. Atmos. Sci., 58, 3224-3242 Collins, W.D., J.K. Hackney, and D.P. Edwards, 2002: A new parameterization for infrared emission and absorption by water vapor in the National Center for Atmospheric Research Community Atmosphere Model”, J. Geophys. Res., 107 (D22), 1-20. Fox-Rabinovitz, M. S., V.M. Krasnopolsky, D.V. Chalikov, 2004: Decadal climate simulations using NN emulations for long wave radiation parameterization, to be submitted Krasnopolsky V., D.V. Chalikov, and H.L. Tolman (2002) A neural network technique to improve computational efficiency of numerical oceanic models, Ocean Modelling, v. 4, 363-383 Krasnopolsky V.M. and F. Chevallier (2003), Some neural network applications in environmental sciences. Part II: Advancing the computational efficiency of environmental numerical models. Neural Networks, 16, 335-348. Krasnopolsky V.M., M.S. Fox-Rabinovitz, and D.V. Chalikov, (2004a): Using neural networks for fast and accurate approximation of the long wave radiation parameterization in the NCAR community atmospheric model: evaluation of computational performance and accuracy of approximation, Proc., 15th Symposium on Global Change and Climate Variations, 84th AMS Annual Meeting, Seattle, Washington, CD-ROM, P1.20 Krasnopolsky V.M., M.S. Fox-Rabinovitz, and D.V. Chalikov, (2004b): Fast and Accurate Approximation of the Long Wave Radiation Parameterization in a GCM Using Neural Networks: Evaluation of Computational Performance and Accuracy of Approximation in the NCAR CAM-2, Proc. CIMSA 2004 – IEEE International Conference on Computational Intelligence for Measurement Systems and Applications, Boston, MD, USA, 14-16 July 2004, CD-ROM, pp. 57-62 Krasnopolsky V.M., M.S. Fox-Rabinovitz, and D.V. Chalikov, (2004c): New Approach to Calculation of Atmospheric Model Physics: Accurate and Fast Neural Network Emulation of Long Wave Radiation in a Climate Model, submitted Krasnopolsky V.M. and M.S. Fox-Rabinovitz, (2004d): A New Synergetic Paradigm in Environmental Numerical Modeling: Hybrid Environmental Numerical Models Consisting of Deterministic and Machine Learning Components, submitted