Canadian Centre for Climate Modelling and Analysis
Modeling Group
Country
Canada
Météo-France / Centre National de Recherches Météorologiques
France
Commonwealth Scientific and Industrial Research Organization in
collaboration with Queensland Climate Change Centre of
Excellence
Geophysical Fluid Dynamics Laboratory
Goddard Institute for Space Studies
USA
Max Planck Institute for Meteorology
Meteorological Research Institute
Germany
Japan
Spatial Resolution
(lon x lat)
3.75° x 3.75°
2.81° x 2.81°
2.81° x 2.81°
Model Name
CCCMA-CGCM3.1(T47)
CCCMA-CGCM3.1(T63)
CNRM-CM 3
ID
C1
C2
D
Australia
CSIRO-Mk3.0
CSIRO-Mk3.5
S1
S2
1.88° x 1.88° *
1.88° x 1.88° *
USA
GFDL-CM2.0
GFDL-CM2.1
GISS-AOM
GISS-EH
GISS-ER
MPI-ECHAM5
MRI-CGCM2.3.2
G1
G2
E1
E2
E3
P
R
2.50° x 2.00° *
2.50° x 2.00° *
4.00° x 3.00°
5.00° x 3.91°
5.00° x 3.91°
1.88° x 1.88° *
2.81° x 2.81°
Table 1. List of the CMIP3 models used for the evaluation of 20th century heavy precipitation and its associated physical mechanisms
in this paper. The approximate spatial resolutions were calculated by dividing 360° or 180° by the number of grid cells in the
longitude or latitude dimensions, respectively.
Asterisks next to spatial resolution denote climate models whose grids were
transformed to the common 2.5°x2.5° resolution using area averaging. All others were transformed using linear interpolation.
Ensemble member run # 1 was used for all models except for the GFDL-CM2.1 and GISS-EH, in which run # 2 and # 5 were used,
respectively. Models were assigned letters (ID column) for ease of reference in analyses showing results from individual models.
[Further documentation for individual models, including expansions of all acronyms, is available online at http://wwwpcmdi.llnl.gov/ipcc/model_documentation/ipcc_model_documentation.php.]
Modeling Group
Beijing Climate Center, China Meteorological Administration
Canadian Centre for Climate Modelling and Analysis
Centre National de Recherches Meteorologiques / Centre Europeen de
Recherche et Formation Avancees en Calcul Scientifique
Commonwealth Scientific and Industrial Research Organization in
collaboration with Queensland Climate Change Centre of
Excellence
LASG, Institute of Atmospheric Physics, Chinese Academy of Sciences
NOAA Geophysical Fluid Dynamics Laboratory
Spatial Resolution
(lon x lat)
2.81° x 2.81°
2.81° x 2.81°
1.41° x 1.41° *
Country
China
Canada
France
Model Name
BCC-CSM1.1
CanESM2
CNRM-CM5
ID
B
C
D
Australia
CSIRO-Mk3.6.0
S
1.88° x 1.88° *
China
USA
FGOALS-s2
GFDL-ESM2G
GFDL-ESM2M
HadGEM2-CC
INM-CM4
IPSL-CM5A-LR
IPSL-CM5A-MR
MIROC-ESM
MIROC-ESM-CHEM
F
G1
G2
H
I
L1
L2
M1
M2
2.81° x 1.67° *
2.50° x 2.00° *
2.50° x 2.00° *
1.88° x 1.25° *
2.00° x 1.50° *
3.75° x 1.88°
2.50° x 1.26° *
2.81° x 2.81°
2.81° x 2.81°
Japan
MIROC5
M3
1.41° x 1.41° *
Germany
MPI-ESM-LR
MPI-ESM-MR
MRI-CGCM3
P1
P2
R
1.88° x 1.88° *
1.88° x 1.88° *
1.13° x 1.13° *
Met Office Hadley Centre
Institute for Numerical Mathematics
Institut Pierre-Simon Laplace
UK
Russia
France
Japan Agency for Marine-Earth Science and Technology, Atmosphere
and Ocean Research Institute (The University of Tokyo), and
National Institute for Environmental Studies
Atmosphere and Ocean Research Institute (The University of Tokyo),
National Institute for Environmental Studies, and Japan Agency
for Marine-Earth Science and Technology
Max Planck Institute for Meteorology
Japan
Meteorological Research Institute
Japan
Table 2. As in Table 1 but for the CMIP5 models used for analysis in this paper. Ensemble member run # 1 was used for all models.
Figure 1. The mean annual precipitation falling from the 99th percentile and above (P99M) during the period 1979-99 for (a) the CPC
observations and the difference between the multi-model average and CPC for (b) CMIP3 and (c) CMIP5.
Figure 2. Scatterplots of RMS error in P99M over North America versus horizontal resolution of CMIP models for (a) annual, (b)
winter, and (c) summer daily data when compared with CPC observations. The abscissa is plotted on a log scale due to the wide range
in spatial resolution between the models. Letters identify the individual models, as defined in Tables 1-2, where red letters correspond
to CMIP3 and blue letters correspond to CMIP5. “A” represents the multi-model average, where the RMS error is computed after first
computing the multi-model average biases at every grid cell. A least squares linear fit to the RMS error versus log(resolution) values
for each model subset and for all the models together is plotted, along with the corresponding correlation coefficients (R2).
Figure 3. Composites of pressure at mean sea level (PRMSL, hPa, contours) and 500 mb
geopotential height standardized anomalies (Z500*, dimensionless, color fills) for the top 1% of
daily annual precipitation events at selected grid cells (indicated by black rectangles, which vary
from top to bottom). Sea level pressure is contoured every 2 hPa. Composites for NARR, the
CMIP3 model average, and CMIP5 model average are shown from left to right, respectively.
Figure 4. The local low-level (10-m to 500-mb mean) winds (m/s, vectors) and vertically
integrated water vapor standardized anomalies (VIWV*, dimensionless, color fills) averaged
over days when precipitation equals or exceeds the annual 99th percentile at every grid cell on
the domain for (a) NARR, (b) the CMIP3 model average, and (c) the CMIP5 model average.
Missing grid cells in the CMIP3 average (b) are the result of missing specific humidity values for
some days in the CSIRO-Mk3.0 model output.
Figure 5. Mean precipitation for the multi-model average of CMIP5 models analyzed in this paper (see Table 2). The mean from the
historical simulations over the period 1979-1999 is shown in panels a through c. The absolute difference between the RCP8.5
simulation period (2079-2099) and historical period is shown in panels d through f, while the percent difference between RCP8.5 and
historical is shown in h through i. The mean over annual, December-February (DJF), and June-August (JJA) days is shown from left
to right, respectively.
Figure 6. As in Fig. 4, but for the mean precipitation falling from the heaviest 1% of daily events (P99M, as defined in the text).
Figure 7. The average error in predictions of cluster assignments for a sample of leftover grid
cells versus number of clusters, based on applying V-fold cross validation to precipitation change
histograms using 5 subsamples (see text for details). The analysis was applied to histograms of
count differences in bins of raw precipitation or precipitation normalized by historical P99M for
either the CMIP5 model average separately or the aggregate of all models (including the model
average), following the legend. In all cases, the log of the count differences in the histograms
was taken before applying the V-fold cross validation analysis. Values on the ordinate axis
represent the sum of squared differences between the predicted centroid and grid cell histograms,
averaged over all grid cells and subsamples left out in the cross validation process.
Figure 8. The (c) cluster assignments and (a,b,d,e) mean histograms over the grid cells assigned to
the respective clusters for the application of k-means clustering to difference histograms of annual
daily precipitation for the model average (see text). The log of the count difference between the
RCP8.5 and historical histograms was taken before applying cluster analysis. The histogram bars
show the mean counts over all grid cells, the box and whiskers show the grid cell inter-quartile
range (25th to 75th percentile, IQR, box) and 5th to 95th percentile (whiskers), and the stars indicate
the minimum and maximum counts of all grid cells for the respective bin and cluster. Precipitation
bin edges are 0, 0.5, 2.5, 5.0, 7.5…125 mm day-1.
Figure 9. As in Fig. 8, but using precipitation normalized by the local historical P99M when
generating histograms. The bin edges are 0, 0.01, 0.1, 0.2, 0.3…4, infinity.
Figure 10. (left) The distance (see text) between the difference histogram at each individual grid
cell and its corresponding assigned centroid histogram, using precipitation normalized by
historical P99M and taking the log of count differences before applying cluster analysis. (right)
Cluster histograms analogous to Fig. 9 but only using grid cells with distances less than 10 units
from their respective centroid histogram to construct the histograms, as opposed to using all grid
cells within the cluster.
Figure 11. As in Fig. 9 but only using grid cells with a distance of 10 or more units from the
respective cluster histogram in Fig. 10.