261 Compound parameterization with a quality control of larger errors for neural network emulations of full model radiation for the NCEP Climate Forecast System model

Monday, 24 January 2011
Washington State Convention Center
Alexei Belochitski, Univ. of Maryland, College Park, MD; and V. M. Krasnopolsky and M. S. Fox-Rabinovitz

Development of neural network (NN) emulations for accurate and fast calculations of physical processes in numerical climate and weather prediction models depends significantly on our ability to generate a representative training set. Owing to the high dimensionality of the NN input vector which is of the order of several hundreds or more, it is rather difficult to cover the entire domain, especially its “far corners” associated with rare events, even when we use model simulated data for the NN training. Moreover, the domain may evolve in time (e.g., due to a climate change). In this situation the emulating NN may be forced to extrapolate beyond its generalization ability and may lead to larger errors in NN outputs.

The compound parameterization (CP) with a quality control (QC) of larger errors, a technique previously developed by the authors to address this problem, has been previously applied to the neural network emulation of short-wave radiation parameterization of the uncoupled moderate resolution NCAR Community Atmospheric Model driven by climatological SSTs. As a first step, CP applies a special NN trained to predict errors in output of NN emulation. Then, if in the given instance the predicted errors are larger then an empirically determined threshold the output of the NN emulation is rejected and the original parameterization's output is used in its place. It is noteworthy that such rejection occurs only for about 1% of profiles. The CP reduces the probability of medium errors by an order of magnitude and practically eliminates large errors (outliers). In this study, the CP with QC is further developed and applied to NN emulations of the full (long- and short-wave) model radiation for the coupled NCEP Climate Forecast System model with significantly higher resolution and time dependent CO2.

ACKNOWLEDGMENTS. The authors would like to thank Drs. S. Lord, Y.-T. Hou, H.-L. Pan, S. Saha, and S. Moorthi for their useful consultations and discussions. The research had been supported by the NOAA CPO CDEP CTB grant NA06OAR4310047.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner