5K differences in Surface Temperature

[liess@umn_edu]

Hi,

We are running CAM3-1.1.p1 on a 10 node dual quadcore 3.0GHz Xeon/16GB RAM system with PGI 7.1-5 and 1 MPI process per compute node and 8 OpenMP threads (OpenMPI 1.2.5) per node.

Validating the first couple of days of our climatological SST runs with the procedure described in http://www.ccsm.ucar.edu/models/atm-cam/port/ results in reasonable results within the specified error range. Also global averages of years 6-15 (we omitted the first 5 years due to spin-up) are similar to the climatology at http://www.ccsm.ucar.edu/models/atm-cam/sims/cam3.0/eul128x256_d50/climo_files/ .

However, when plotting global maps for each month, we see differences of 5K in Surface Temperature and Reference Height Temperature with one negative peak over Siberia and a positive peak over Alaska during DJF and a similar pattern over the same longitudes over the Antarctic in JJA when comparing our results to the monthly mean climo_files mentioned above. Should we concerned about this or is it part of model uncertainties, when running on different architectures? The radiation budget (FSNT and FSNTOA) shows a few differences over 20 W/m**2, but looks generally OK. We are especially interested in results over the Arctic.

Thanks in advance and best regards,
Stefan

 

[rneale]

When you say differences for each month, do you mean a monthly climatology or an individual month? And do you mean the 5K difference is between the surface temperature and the surface air temperature or differences with the NCAR control data for each of these fields?

 

[liess@umn_edu]

We actually tried different strategies, e.g. using only individual months of the first year, and an arbitrary year, as well as 5-year mean climatologies. Our results were relatively consistent to each other compared to the monthly mean files available at ucar's website, which show these differences. We are wondering if either the ucar files are not produced with the same CAM3-1.1.p1 version we use, or if we might have made an error in our model compilation or in the choice of forcing data.

 

[rneale]

There is no easy similarity test for climate simulations such as you are doing. Inevitably they will be different because of different machines and slightly different compiler options, but if we see significant differences on a large scale with a 10-year climatology then we get concerned. Do you have an example plot I could look at?

 

[rneale]

The tropical signals are somewhat concerning. If we are nervous about differences then we here tend to run for longer (10-15 years) to see if the differences get smaller, particular over the arctic which seems the largest interannual variability.