Sensitivity tests are presented which demonstrate how global climate models (GCMs) respond to longwave radiation input. This input is established by the inclusion in the GCM of a longwave radiation heating algorithm, and a number of such algorithms are tested in the NCAR/CCM2.

Climate model dynamics are driven by forcing and the primary forces affecting the thermal field are long wave radiative heating (LWR), short wave radiative heating (SWR), and convection (cumulus, etc.). These forcing effects are cycled through the thermal field to the motion field by nonlinear transfer. The model dependent variables, in particular temperature (T), moisture (Q) and especially clouds, evolve in time in a GCM and determine the subsequent forcing. If the dependent variables are not accurately calculated in space and time, the predicted forcing functions will be adversely affected. As integration time proceeds, such inaccuracies will lead to systematic errors in the prediction of climate. Although the effects of all these forces need evaluation to understand their impact on climate, we focus here on one of the primary forces, the long wave radiative heating.

This forcing is determined in GCMs by a LWR heating algorithm. Such algorithms compute vertical heating rate (HR) profiles from profiles of T, Q, clouds, and minor constituents. The HRs thus depend on the vertical structure of the input variables plus internal physics. To test the sensitivity to input variables of various algorithms taken from global climate models, we intercompared the heating rate profiles produced by six GCM algorithms each receiving identical input conditions from two distinct weather situations. One had clear skies and the other had deep clouds. The algorithms proved to be extremely sensitive to clouds, yielding remarkably different heating rates for identical input conditions. For more details, see Baer, et al., 1996.

Noting this sensitivity, we must determine how global climate models (GCMs) respond to the HRs generated by the LWR algorithms. The LWR algorithms calculate vertical profiles of HRs periodically at each point in a GCM generating 3-dimensional fields; these HR fields represent the LWR forcing which determines the temperature-tendencies. The temperature field predicted from these tendencies modifies the wind field tendencies by nonlinear interaction in space. The predicted wind field then subsequently modifies the temperature tendencies by nonlinear advection. Thus the impact of the HRs is spread in time and space to all the variables (T, Q, clouds), and these modified variables are then used to recalculate new Hrs. For details on modeling climate see Washington and Parkinson, 1986.

We choose to present the model output in spectral statistics as well as in physical space. This is provoked because climate model statistics in physical space generally require many realizations and computer time is very limited. Thus if we can reduce the number of realizations from which we can derive meaningful statistics, we will be able to achieve more experiments. Spectral statistics provide scaling information and allow for separation of amplitude from phase. The variability in phase seems much more pronounced than that in amplitude. If so, perhaps only a few realizations may yield a good representation of the amplitude. We tested this hypothesis with data from a long run with a two level baroclinic model given 80 realizations, and indications from this experiment showed that amplitudes did indeed vary considerably less than phase, thus suggesting that even an individual realization may provide useful information.

Three dimensional fields from model output archives were utilized to demonstrate sensitivity of GCMs to longwave radiative heating. Model output of LWR HRs and other fields at each archived time were averaged over 60 day wintertime (Jan-Feb.) periods to develop climate statistics. The model archives available for analysis included;

• NCAR CCM1: R15, T42; 12 levels; SST climatology
• NCAR CCM2: R15, T42, T106; 18 levels; SST climatology
• NCAR CCM2: T42; 18 levels; AMIP SSTs
• CSU: 4x5x10 levels ; AMIP SSTs
• NASA GEOS-1: 4x5x20 levels ; AMIP SSTs

Unfortunately, no observations are available for comparison. We therefore used a model to generate clouds from observed (T, Q) data, and used those clouds to generate HRs in the model. We introduced the observed data into the CCM2 as initial conditions for each day of the AMIP period (J-F, 1987) and ran the model for 36 hours. For information on AMIP, see Gates, 1992. We used the HR fields which the model developed at that time and defined them as "pseudo-observations" to compare with the output from various AMIP model runs. For comparison, we repeated the above calculation with the CCM2 but substituted the ECMWF LWRM algorithm (see Morcrette, 1990).

The results of the intercomparisons of model output showed the following sensitivities.

• Model truncation-- Analysis of both the CCM1 and CCM2 climatology archived data showed that HRs are much more dependent on model truncation in the CCM1 than the CCM2.
• Model changes, including changes to the incorporated LWR algorithm-- HRs are strongly dependent on model changes as observed when comparing output from the CCM1 and CCM2 climatology runs and also from comparison of archived AMIP data from the CCM2, CSU and GEOS-1 GCM runs.
• Surface forcing in the form of sea-surface temperature (SST)-- This forcing appears to have a significant effect on the evolution of the model variables as seen when comparing two CCM2 AMIP periods (J-F, 87 and J-F, 83) and the winter months from the CCM2 climatology run .
• Control integrations/ pseudo-observations-- When comparing a CCM2 AMIP winter with our pseudo-observations using both the CCM and ECMWF algorithms, the model output fields of the pseudo- observations calculation compare more favorably to one another than to the AMIP run output. This indicates strong dependence on the initial data entering the LWR algorithm.

To test the effect of longwave radiative forcing with other GCM features held fixed, we ran the CCM2 with two different LWRM algorithms, the NCAR version native to the model and the ECMWF (Morcrette) algorithm implanted. Both integrations are for two months of the AMIP period, Jan.-Feb./87, with identical initial conditions. From averaged heating rates for this experiment on the 250hPa surface compared to the corresponding pseudo- observation maps, we note differences on the order of 10 percent, which is a significant difference, and the ECMWF algorithm yields stronger cooling. To determine if these differences are due to the algorithms or to climate variability, we must establish the GCMs inherent variability.

Any one realization of climate statistics may not reflect the climate for that period. To determine the effect of climatic variability, we have run ten realizations of the CCM2 with only slight variations in the initial states. We began the runs at the beginning of October 1986, allowed two months for equilibrium, and took model output statistics from the subsequent three months of DJF. The ten runs used the NCAR radiation algorithm (see Kiehl et al., 1994). Means of the heating rates from the ten realizations and their standard deviations were taken from the integrations and 200 and 500hPa maps were created. These maps showed substantial variability in the runs with standard deviations at many locations in excess of ten percent. Similar results were found in the fields of clouds and temperature.

To determine if the variability between model runs with different heating algorithms exceeds the climate variability of a model with no algorithm change (results just described), we ran three realizations with the CCM2 including the ECMWF algorithm and then repeated the same test using the NCEP algorithm. Testing for climate variability with three realizations appears realistic. Comparing averages of the three realizations with the averages of the ten CCM realizations, we are able to demonstrate the impact of algorithm change. On Figure 1 we show the difference maps of the average temperature at both 200 and 500hPa, comparing the ten CCM runs with the three runs each using the ECMWF, NCEP and CCM algorithms. All three runs for each case were for the same initial conditions. The two right panels used the same algorithm; thus their difference reflects model climate variability. The other panels reflect variability dependent on algorithm choice. It is evident that the climate statistics for the runs with different algorithms show much larger variability from one another than is seen for the model climate variability. This effect is also evident if a scale dependent diagram for heating rates is developed.

Figure 1. Model output statistics of temperature fields averaged over DJF (1986-87) using different algorithms in the CCM2. Data from 3 realizations of each run using the 3 algorithms (CCM, ECMWF, NMC) were compared by differencing to the CCM2 output for 10 realizations.

Figure 2. Temperature fields at 200hPa averaged over DJF (1986-87) using different algorithms in the CCM2 and averaged over 3 realizations. Output using all 3 algorithms is presented as well as differences between the 3 maps.

Having established the significance of climate statistics for runs with different LWR algorithms, we now compare integrations using the CCM, ECMWF and NCEP algorithms. Since we have only three realizations using the ECMWF and NCEP algorithms, we compare the average fields from those three realizations with the same three realizations using the CCM algorithm. We have considered the 500hPa mean fields of heating rates for the three experiments together with their absolute differences. The same analysis has been performed for temperature on the 200hPa surface and is presented on Figure 2. One can see immediately from this figure that the CCM and ECMWF runs compare more favorably to one another than to the NCEP(NMC) run. Additionally, the differences on all maps exceed ten percent in many locations on the globe, differences which are far above the climate variability of the model.

Using different LWRM algorithms in the CCM2 results in notable differences of model output when the integration proceeds for 60 days. This contrasts to the similarities found from integrations after 36 hours. From climate validation studies as discussed, the impact of HR algorithms in a GCM shows significant climate variability. The climate statistics developed using three different LWR algorithms in a GCM (the CCM2) when compared show much greater variability than the climate variability of the GCM itself. Of the three algorithms tested, the NCEP algorithm showed greater differences from the other two, the NCAR and ECMWF algorithms. Further tests are needed with different models. Plans are underway to perform the same experiment described here with the ECMWF, NCEP and NASA GCM models.

Baer, F., N. Arsky, J. J. Charney, and R. G. Ellingson, 1996: Intercomparison of heating rates generated by global climate model longwave radiation codes. J. Geophys. Res., accepted for publication.

Gates, W.L., 1992: AMIP: The Atmospheric Model Intercomparison Project. Bull. Amer. Meteor. Soc., 73, 1962-1970.

Kiehl, J.T., J.J. Hack, and B.P. Briegleb, 1994: Simulated Earth radiation budget of the National Center for Atmospheric Research Community Climate Model 2 and comparisons with the ERB experiment. J. Geophys. Res., 99, 20815-20827.

Morcrette, J.-J., 1990: Impact of changes to the radiation transfer parameterizations plus cloud optical properties in the ECMWF model. Mon. Weather Rev., 118, 847-873.

Washington, W.M., and C.L. Parkinson, 1986: An Introduction to Three-Dimensional Climate Modeling, 422 pp., Univ. Sci. Books, Mill Valley, Calif.