US National Oceanic and Atmospheric Administration
V. M. Krasnopolsky
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||

Calculation of model physics in a GCM (General Circulation Model) usually takes a very significant part of the total model computations. Evidently, this percentage is model dependent but full model radiation is the most time-consuming component of GCMs ( For example, in the majority of modern radiative schemes, the correlated-k method (Lacis and Oinas 1991) is typically used to reduce the integration over wavelength by effectively binning wavelengths with similar absorption coefficients (k-terms). This simplification reduces greatly the number of monochromatic radiative transfer calculations required. The number of k-terms can be adjusted, which provides a trade-off between the accuracy and efficiency required for a given application. However, the correlated-k methods cannot be made sufficiently computationally efficient to allow calculations for every grid-point at every time-step. To reduce the cost further, calculations are usually made at lower temporal and/or spatial resolutions. Quite drastic reductions in temporal resolution are often made ( Such a situation is an important motivation for developing new alternative numerical algorithms that provide faster calculations of model physics while carefully preserving their accuracy. In our previous studies (Krasnopolsky
It this study we applied the NN emulation approach to the higher complexity NCEP CFS (Climate Forecasting System), which required further development of the neural network emulation methodology. We demonstrate that the NN emulation approach for model radiation can be successfully applied to the
In Section 2, the improved NN emulation approach and developed NN emulations for the NCEP CFS long-wave radiation (LWR) and short-wave radiation (SWR) are briefly described in terms of their design, accuracy, and computational performance. In Section 3, the results of the parallel decadal model simulations, one using both LWR and SWR NN emulations for calculation of full model radiation and the other using the original model radiation (the control run) are compared in terms of similarity of their spatial and temporal variability characteristics. Section 4 contains conclusions.
NN emulations of model physics are based on the two following facts. First, any parameterization of model physics is a continuous or almost continuous mapping (input vector vs. output vector dependence) and can be symbolically written as: (1) where M denotes the mapping, n is the dimensionality of the input space, and m is the dimensionality of the output space. And second, NNs (multilayer perceptrons) are generic tools for approximation of such mappings (Funahashi 1989). NN is an analytical approximation that uses a family of functions like: (2) where yare components of the input and output vectors
_{q }X and Y, respectively, a and b are fitting
parameters, and
is a “neuron”. The activation function
Φ is usually a hyperbolic tangent, n and m
are the numbers of inputs and outputs (the same n and m as
in Eq. (1)), respectively, and k is the number of neurons in the
hidden layer. Definitions of NN terminology can be found in many
places, for example in the recent book by Bishop (2006) and in the
review paper by Krasnopolsky (2007); however, eq. (2) is sufficient to
understand the subject of this paper. The numerical complexity of NN
(2) can be well approximated by a number of NN weights (Krasnopolsky
2007):
The NN
numerical complexity T, required
for the estimating NN (2),_{NN}
N with the coefficient of
proportionality _{C}c depending mainly on a hardware and software
environment of the computer used.
The LWR and SWR parameterizations together comprise the full model radiation. The LWR and SWR parameterizations or the full model radiation for the NCEP CFS have been emulated using two NNs, one for LWR and another for SWR. The input and output vectors for NNs, emulating the LWR or SWR parameterizations, include the same parameters as those of the input and output vectors for the original LWR or SWR parameterizations, respectively. For the LWR NN emulation, these input parameters are the following nine profiles: atmospheric pressure, temperature, specific humidity, ozone mixing ratio, total cloud fraction, cloud liquid water path, mean effective radius for liquid cloud, cloud ice water path, and mean effective radius for ice cloud. The LWR parameterization (and LWR NN emulation) output vectors consist of the profile of heating rates (HRs) and five radiation fluxes: the total sky outgoing LW radiation flux from the top layer of the model atmosphere (the outgoing LWR or OLR), the clear sky upward flux at the top of the model atmosphere, the total sky upward flux at the surface, the total sky downward flux at the surface, and the clear sky downward flux at the surface.
The NN emulation of the LWR parameterization includes
all non-constant inputs of the original LWR (total 556 inputs; n = 556
in eq. (1)). It has the same outputs (total 69 outputs; m = 69 in eq.
(1)) as the original LWR parameterization. We have developed several NNs,
all of which have the same aforementioned inputs and outputs, with the
number k changing from 50 to 200 in eq. (2). Varying k, the number of
terms (or neurons) in eq. (2), allows us to demonstrate the dependence
of the accuracy of approximation on this parameter as well as its
convergence, and as a result, to provide a sufficient accuracy of
approximation for the model (e. g. Krasnopolsky The input vectors for the SWR parameterization include 55 vertical profiles: atmospheric pressure, temperature, specific humidity, ozone, CO2, N2O, O2, and CH4 volume mixing ratios, total cloud fraction, cloud liquid water path, mean effective radius for liquid cloud, cloud ice water path, mean effective radius for ice cloud, and three profiles (optical depth, single scattering albedo, and asymmetry parameter) for each of 14 different species of aerosols. The input vectors include also the solar zenith angle, the solar constant and the surface albedo for four different bands. The SWR parameterization output vectors consist of a vertical profile of heating rates (HRs) and nine radiation fluxes: three fluxes at the top layer of the model atmosphere (the total sky outgoing SW radiation flux, the total sky downward flux, the clear sky upward flux), four radiation fluxes at the surface (the total sky upward and downward fluxes and the clear sky upward and downward fluxes), and the downward (the total and clear sky) fluxes in the UV-B spectral band. The NN emulations of the SWR parameterization have 562 inputs and 73 outputs. We have developed several NNs, with the number k changing from 50 to 200 in eq. (2). It is noteworthy that, as in the case of the NN emulation of LWR, the number of NN inputs is less than the number of input profiles multiplied by the number of vertical layers plus the number of relevant single level characteristics. Many input variables (e.g., almost all gases) have zero or constant values for the upper vertical layers, and for some gases the entire volume mixing ratio profile is a constant (obtained from climatological data).
The NCEP CFS (T126L64) has been run for seventeen
years to generate representative data sets. The representative data set
samples adequately the atmospheric state variability, i.e., it
represents all possible states produced by the model as fully as
possible (including the states introduced due to time dependent CO2
concentration). All inputs and outputs of original LWR and SWR
parameterizations have been saved for two days per month, It is noteworthy that along with the aforementioned requirement of representing all possible states produced by the model, the size of the training data set is limited mainly by the training time, which, in turn, is determined by the processor type and the amount of memory available. The training time is approximately proportional to the size of the training data set. In our case, the selection of about 200,000 input/output records for training is a result of an optimal choice providing a sufficient representativeness of the training set and a reasonable training time. We selected the size of the test set equal to the size of the training set because the training and test sets are supposed to have close statistical properties. There are no serious limitations to the size of the validation set; we selected it equal to the size of the first two sets.
To ensure a high quality of representation of the LWR and SWR processes, the accuracy of their NN emulations has been carefully investigated. The NN emulations have been validated against the original NCEP CFS LWR and SWR parameterizations. To calculate the error statistics presented in Table 1; the original parameterizations and their NN emulations have been applied to the validation data set. Two sets of the corresponding HR profiles have been generated for both LWR and SWR. Total and level bias (or a mean error), total and level RMSE, profile RMSE or PRMSE, and
N (3) and speedup _{C}η (how many times NN
emulation is faster than the original parameterization) are shown.
RRTMG and RRTMF are different versions of the radiation code developed
by AER Inc.Using NN emulations simultaneously for LWR and SWR or for the full model radiation results in an overall significant, about 20 – 25% speedup of NCEP CFS climate simulations when both LWR and SWR are calculated every hour. The speedup η provided by NN emulations (see Table 1) can be also used for more frequent calculations of model radiations. For example, for calculations with higher (T382) model horizontal resolution, if full NN radiation is calculated 10 times more frequently, i.e., every six minutes, at every model dynamics time step (instead of every hour), the time required for the climate simulation using full NN radiation will be still less than the time needed for the climate simulation using the original radiation with the one hour frequency.
In this section we present comparisons between two parallel 17-year NCEP CFS model runs: one using the original LWR and SWR (the control run) and another one using their NN emulations. Both spatial and temporal characteristics of prognostic and diagnostic fields are compared for the parallel runs. To better estimate the changes introduced by NN emulations, we compare them with “background changes” between two control runs performed with the original NCEP CFS model configuration, i.e., without NN emulations. The first run was performed before and the second run after the routine changes (introduced quasi-regularly by system administrators) of the version of the FORTRAN compiler and libraries. The results of 17-year climate simulations performed with NN emulations for both LWR and SWR, i.e., for the full model radiation, have been validated against the parallel control NCEP CFS simulation using the original LWR and SWR. We analyze the differences between the parallel runs in terms of time and spatial (global) means as well as temporal characteristics. |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||

Let us discuss first the differences between the parallel simulations in terms of spatial and temporal radiation characteristics. The differences between the NN radiation and control runs and the differences between two control runs for zonal and time mean LWR and SWR fluxes are presented in Fig. 1. For fluxes presented in Fig. 1 both the differences between the NN radiation and control runs and the differences between two control runs are small and similar by magnitude. They do not exceed 2-3 W/m Let us discuss now prognostic and diagnostic characteristics such as SST, precipitation, different types of clouds, and time series that are sensitive to changes in the model resulted from using NN emulations. Close similarities have also been obtained for these results of parallel runs in terms of time mean spatial fields, which are presented in Figs. 2 to 4. Figs. 2 to 4 have the same design: the upper left panel shows fields produced in the control run (CTL) and the upper right – in the full radiation NN run. The bottom left panel shows the difference (bias) between the full radiation NN and CTL runs, and the bottom right panel shows for comparison the background differences (between two control runs) described above. The 17-year (1990-2006) time-mean SST distributions and bias/differences for the full radiation NN run vs. the control run and the background differences between two control runs are presented for summer in Fig. 2. The SST bias is very small; it is not larger than the background differences. The results for other seasons are similar. Figure 3 shows the 17-year (1990-2006) time-mean distributions and bias/differences for total precipitation (PRATE) for the parallel full radiation NN and control runs for summer. The PRATE bias is quite limited and occurs mostly in the tropics; it is also very close by magnitude and pattern wise to the background differences. The results for other seasons are similar. Figure 4 shows comparisons for the parallel full radiation NN and control runs for total clouds, which are very close for the above runs. The results for other seasons and for other types of clouds are similar.
In this study, the NN emulation approach (Krasnopolsky et al. 2005, 2008) is implemented in the state-of-the-art, high resolution, coupled NCEP CFS. The developed highly accurate neural network emulations of long-wave (RRTMG-LW and RRTMF-LW) and short-wave (RRTMG-SW) radiation parameterizations are 12 to 45 times faster than the original/control long-wave and short-wave radiation parameterizations, respectively. The use of the full NN model radiation results in: (1) an overall speedup of about 20 – 25% for climate simulations and seasonal predictions, and (2) an opportunity to increase significantly the frequency of radiation calculations (for example, to calculate model radiation at every model dynamic time step) without increasing the total model calculation time. |
Figure 1
Figure 2
Figure 3
Figure 4 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||

Bishop, Ch. M., 2006: Pattern Recognition and Machine Learning.
Côté,
J., S. Gravel, A. Méthot, A. Patoine, M. Roch, and A. Staniforth, 1998:
The operational CMC-MRB global environmental multiscale (GEM) model.
Part I: Design considerations and formulation,
Funahashi, K., 1989: On the Approximate Realization of Continuous
Mappings by Neural Networks.
Kalnay, E.,
Krasnopolsky, V.M., 2007: Neural Network Emulations for Complex
Multidimensional Geophysical Mappings: Applications of Neural Network
Techniques to Atmospheric and Oceanic Satellite Retrievals and Numerical
Modeling,
Krasnopolsky, V.M., M.S. Fox-Rabinovitz, and D.V. Chalikov, 2005: Fast
and Accurate Neural Network Approximation of Long Wave Radiation in a
Climate Model,
Krasnopolsky, V.M., M.S. Fox-Rabinovitz, A. A. Belochitski, 2008:
"Decadal Climate Simulations Using Accurate and Fast Neural Network
Emulation of Full, Long- and Short Wave, Radiation.", Krasnopolsky, V.M., M. S. Fox-Rabinovitz, Y. T. Hou, S. J. Lord, and A. A. Belochitski, 2009: Accurate and Fast Neural Network Emulations of Model Radiation for the NCEP Coupled Climate Forecast System: Climate Simulations and Seasonal Predictions, submitted. Lacis, A.
A., V. Oinas, 1991: A description of the correlated k-distribution
method for modeling non-gray gaseous absorption, thermal emission and
multiple scattering in vertically inhomogeneous atmospheres. Manners,
J., J.-C. Thelen, J. Petch, P. Hill & J.M. Edwards, 2008: Two fast
radiative transfer methods to improve the temporal sampling of clouds in
NWP and climate models, Morcrette,
J.-J., G. Mozdzynski and M. Leutbecher, 2008: A reduced radiation grid
for the ECMWF Integrated Forecasting System, |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||

Contact
V. M. Krasnopolsky |