Fast Nn Emulation of the Super-parameterization in the Multi-scale Modeling Framework

FAST NN EMULATION OF THE SUPER-PARAMETERIZATION IN THE MULTI-SCALE MODELING FRAMEWORK

 

V. M. Krasnopolsky NOAA/NCEP/EMC College Park, MD, M. S. Fox-Rabinovitz University of Maryland, College Park, MD, P. J. Rasch, and M. Wang, PNNL, Richland, WA

 

Extension of the NN emulation framework for super-parameterization (SP) has been developed.  The approach, originally labeled as a “super-parameterization” (SP), but later as the Multi-scale Modeling Framework (MMF), is attracting a growing interest in the climate modeling community because it couples cloud-scale and large-scale dynamics within a single modeling framework. However, this approach increases the model run time by a factor of 200 to 250, which severely limits its applicability. To address this problem we propose to develop accurate and very fast neural network emulations of SP. The current PNNL-MMF (Wang et al. 2011) has been used to provide the training set for a NN emulator (Krasnopolsky 2013). The large scale fields from the MMF that provide the SP inputs are used as the NN input vector and the aggregated results of the SP are used as the output vector. Multiple NNs with n = 746 inputs, m = 121 outputs, and k = 25 to 150 hidden nodes have been trained. The limited training dataset was drawn selectively from four days of MMF global simulations, spanning diurnal and latitudinal variability.

Fig. 1 Taylor diagram for all seven SP outputs of all ten ensemble member NNs together with the Ensemble Mean (EM).  Seven different outputs and their EMs are presented with different symbols of different colors (see the figure legend). Solid isolines identify skill score values, dashed lines indicate correlation. The number of the ensemble member indicated by a number written beside the symbol.

 

Experiments with different NN architectures have been performed and assessed.  These experiments include comparisons of a single emulating NN with many outputs vs. a battery of simpler emulating NNs with different outputs.  Also experiments with different output normalizations and different emulating NN complexities have been conducted. Training an ensemble of emulating NNs or developing NN-SP has been performed.  Performance (accuracy and speedup) of the NN-SP has been estimated on an independent simulated dataset.  Multiple statistics have been calculated for NN-SP and the differences between SP and NN-SP.  Fig. 1 shows Taylor diagram for NN-SP outputs.

 

The results obtained in this work can be summarized as follows: (1) The SP can be emulated by NN with a satisfactory accuracy (based on validation on an independent data set). Whereas the accuracy of NN-SP for some small-scale variables (e.g., spdqc - shown by green circles and spdqi - shown by blue triangles) should be improved, other larger-scale variables are emulated with NN-SP with rather high accuracy.; (2) The NN-SP provides an impressive, two orders of magnitude, speedup as compared with the original SP. It provides a practical opportunity to use NN-MMF for decadal and longer climate simulations.

 

References

Krasnopolsky, V.M., 2013: "The Application of Neural Networks in the Earth System Sciences. Neural Network Emulations for Complex Multidimensional Mappings", 200pp., Springer

Wang, M., S. Ghan, R. Easter, M. Ovchinnikov, X. Liu, E. Kassianov, Y. Qian, W. I. Gustafson, V. E. Larson, D. P. Schanen, M. Khairoutdinov, and H. Morrison, 2011: The multi-scale aerosol-climate model PNNL-MMF: model description and evaluation, Geosci Model Dev, 4(1), 137-168.