Introduction
The mass balance of the Antarctic Ice Sheet is a major source of uncertainty in
estimates of projected sea-level rise. Projections of Antarctic mass changes
are based on the input–output method, in which ice-sheet surface mass balance
(SMB, input) and ice-sheet dynamics (output) are modelled separately. The mass
budget of the Antarctic Ice Sheet is 10 times lower in magnitude than the
individual input/output components. Consequently, when using the input–output
method, uncertainty in the total mass budget equals the sum of the
uncertainties of input and output estimates, which are of the same order of
magnitude as the mass budget itself. This drives efforts to better estimate and
reduce uncertainty on each of these two components.
The SMB of the Antarctic Ice Sheet is driven by snowfall at the ice-sheet
margins, although sublimation, melt, refreezing, and drifting snow can be of
importance locally. These components cannot be reliably deduced from reanalyses
or global climate models (GCMs) because their horizontal resolution
(∼100 km) is insufficient and because their physics are not
adapted for cold and snow-covered regions. Polar-oriented regional climate
models (RCMs) are able to fill this gap because their physics have been
specifically developed/calibrated for these areas. Forced with reanalyses,
their results can be evaluated directly against meteorological, remote-sensing
and SMB observations available in these high-latitude regions. With regard to
climate change, the response of the cryosphere will depend both on its initial
state and on the climate change signal. Accordingly, RCM results will rely on
the ability of GCMs to adequately simulate the current climate as well as on
GCM estimates of future changes.
Unlike previously published evaluations of the Coupled Model Intercomparison
Project Phase 5 (CMIP5) models over Antarctica
which focus on specific fields such as westerly winds
or sea ice
, in this paper we aim to
evaluate the CMIP5 fields that will be used as input for RCMs (atmospheric
fields at lateral boundaries and surface oceanic conditions into the
integration domain) and those that may have the greatest impact on RCM-based
SMB components (air temperature, air humidity, surface pressure, sea-ice
concentration and sea surface temperature).
After describing models, measures and variable selection in
Sect. , we perform multi-variable analysis and
establish relationships between climate change in GCMs and their representation
of the current climate in Sect. . We conclude by discussing
potential sources of bias in our method and by summarizing our main outcomes.
Data and methods
CMIP5 climate models and reanalyses
Monthly mean fields from 41 CMIP5 models and 6 reanalyses, listed in
Table , are compared in this work. All data were bi-linearly
interpolated onto a common regular longitude–latitude horizontal grid
(1.5∘ × 1.5 ∘) with a spatial domain extending south of
40∘ S over the ocean. We did not include land and ice-covered areas
because (i) RCM lateral boundaries are set over the ocean when possible and
(ii) RCMs are never forced by GCM outputs over the land surface, except for
the initialization. Seasonal values are defined by 3-month means, with
winter consisting of June-July-August for atmospheric variables and
July-August-September for oceanic variables. All other seasons are defined
with a similar 1-month lag for oceanic variables.
CMIP5 data were retrieved from the historical (1850–2005 period) and
representative concentration pathway 8.5 (“RCP85”, 2006–2100 period) coupled
ocean–atmosphere experiments. The RCP85 scenario is an upper range of plausible
future emission in which greenhouse gas radiative forcing continues to rise
throughout the 21st century until the 1370 ppm CO2 equivalent
. In this scenario, stratospheric ozone recovery is
represented across the CMIP5 models, with recovery over Antarctica to near
pre-ozone hole amounts by 2100. We merged historical and RCP85 to form
continuous time series from 1850 to 2100. We focused on the first realization
(r1i1p1) but also considered r2i1p1 and r3i1p1 realizations, when available,
to check the robustness of our results. Given the high number of models
investigated, we highlighted models which contained obvious similarities in
code or were produced by the same institution (colours in
Figs. and ), following the work of
colours in their Fig. 1.
Recent reanalysis inter-comparisons have shown the European Centre for
Medium-Range Weather Forecasts “Interim” re-analysis ERA-Interim,
1979–present; to be the most reliable contemporary global
reanalysis over Antarctica ,
prompting our choice of ERA-Interim as a reference for representing the current
climate (1980–2010). However, comparisons with five other reanalyses were also
performed in our study: the Japanese 55-year Reanalysis from the Japan
Meteorological Agency JRA-55, 1958–present;, the National
Aeronautics and Space Administration Modern-Era Retrospective Analysis for
Research and Applications MERRA, 1979–present;;
the National Centers for Environmental Prediction (NCEP)/National Center for
Atmospheric Research Global Reanalysis 1 NCEP-NCAR-v1,
1948–present;; the NCEP/Department of Energy Atmospheric
Model Intercomparison Project 2 reanalysis NCEP-DOE-v2,
1979–present;; and the National Oceanic and Atmospheric
Administration (NOAA) Twentieth Century Reanalysis v2 NOAA-20CR-v2,
1870–2012,.
We will later define measures to compare CMIP5 GCM outputs with ERA-Interim
over the period 1980–2010 (31 years). In order to reduce the sensitivity of
our comparisons to the choice of this reference period, we computed the
multi-decadal intrinsic variability of those measures. Over the Antarctic
region considered, CMIP5 GCM metrics show no significant trends until the
1980s, but they evolve significantly afterwards. Consequently, we estimated the
multi-decadal climate variability of each metric for every CMIP5 GCM by
considering the variability of the 31-year running metric during the
stable period 1850–1980. We present this estimate in detail in Appendix
. The multi-decadal
variability estimate gives an error bar around the reference period value,
which depends on each metric and each model (Table ).
Reanalyses (first six rows) and CMIP5 model details. Climate prediction
indexes (CPIs) are given plus/minus estimate of the multi-decadal variability.
Ranks given in parentheses are the modified ranks when using CPI plus/minus
multi-decadal variability for the considered model while not changing CPIs of
other models. On the ERA-Interim line, we give the ERA-Interim standard
deviation of spatially averaged annual values, which are the scaling factors
for the indexes, and when combining several seasons we give the mean standard
deviation plus/minus (maximum–minimum)/2.
Name
Modelling
Lat.
CPI and ranks
groups
grid
msie[win]
prw[s/w]
psl[ann]
ta850[s/w]
ta850[sum]
tos[sum]
spacing
CPI
Rank
CPI
Rank
CPI
Rank
CPI
Rank
CPI
Rank
CPI
Rank
ERA-Interim
ECMWF
0.7∘
–
0.75 ± 0.1 kgm-2
3.2 ± 0.5 hPa
0.95 ± 0.06 K
0.89 K
0.56 K
JRA-55
JMA
1.25∘
0.5 ± 1.0
4 (2–5)
0.6 ± 0.5
3 (2–10)
0.2 ± 0.4
4 (2–9)
0.7 ± 0.4
3 (3–10)
0.8 ± 0.4
5 (3–11)
0.9 ± 0.9
6 (2–7)
MERRA-v1
NASA
0.5∘
0.1 ± 1.0
2 (2–5)
0.5 ± 0.5
2 (2–5)
0.1 ± 0.4
2 (2–6)
0.3 ± 0.4
2 (2–2)
0.3 ± 0.4
2 (2–2)
0.2 ± 0.9
2 (2–6)
NCEP-DOE-v2
NCEP-DOE
2.5∘
0.4 ± 1.0
3 (2–5)
2.5 ± 0.5
40 (37–42)
0.3 ± 0.4
5 (2–10)
1.0 ± 0.4
7 (3–23)
0.9 ± 0.4
7 (3–14)
0.4 ± 0.9
4 (2–6)
NCEP-NCAR-v1
NCEP-NCAR
2.5∘
0.5 ± 1.0
5 (2–6)
2.0 ± 0.5
36 (28–39)
0.2 ± 0.4
3 (2–7)
0.8 ± 0.4
4 (3–14)
0.7 ± 0.4
3 (3–11)
0.3 ± 0.9
3 (2–6)
NOAA-20CR-v2
NOAA
2.0∘
3.6 ± 1.0
29 (23–38)
1.9 ± 0.5
31 (21–37)
0.3 ± 0.4
6 (2–14)
1.0 ± 0.4
6 (3–23)
0.9 ± 0.4
6 (3–13)
0.6 ± 0.9
5 (2-6)
ACCESS1-0
CSIRO-BOM
1.25∘
1.9 ± 0.4
11 (6–17)
1.0 ± 0.3
7 (4–16)
0.6 ± 0.2
9 (7–21)
1.1 ± 0.1
9 (6–11)
1.3 ± 0.1
15 (12-15)
3.7 ± 0.4
28 (25–32)
ACCESS1-3
CSIRO-BOM
1.25∘
2.1 ± 0.2
15 (12–18)
1.1 ± 0.2
8 (5–15)
0.7 ± 0.2
10 (7–22)
0.9 ± 0.2
5 (3–8)
0.8 ± 0.2
4 (3–7)
2.7 ± 0.3
14 (11–22)
BCC-CSM1-1
BCC
2.8∘
3.1 ± 0.5
28 (23–29)
1.9 ± 0.3
33 (28–37)
1.3 ± 0.2
35 (35–37)
1.2 ± 0.3
12 (6–27)
1.1 ± 0.3
11 (6–15)
2.1 ± 0.4
8 (7–13)
BCC-CSM1-1-m
BCC
1.0∘
4.0 ± 1.5
31 (20–42)
1.9 ± 0.4
34 (26–37)
1.4 ± 0.1
37 (35–37)
1.1 ± 0.3
8 (4–22)
1.0 ± 0.3
9 (3–15)
2.2 ± 0.5
10 (7–15)
BNU-ESM
GCESS
2.8∘
6.7 ± 0.9
46 (45–47)
2.0 ± 0.4
35 (28–39)
1.8 ± 0.3
41 (38–47)
2.3 ± 0.4
44 (38–45)
1.5 ± 0.3
19 (12–31)
3.3 ± 0.4
26 (18–28)
CanESM2
CCCma
2.8∘
2.1 ± 0.5
14 (8–22)
1.3 ± 0.4
18 (6–30)
0.7 ± 0.2
15 (8–26)
1.9 ± 0.4
37 (28–44)
1.8 ± 0.4
31 (16–38)
2.2 ± 0.3
9 (8–10)
CCSM4
NSF-DOE-NCAR
1.25∘
2.7 ± 0.5
23 (16–28)
1.3 ± 0.1
17 (12–20)
1.0 ± 0.2
28 (14–34)
1.2 ± 0.4
13 (5–29)
1.1 ± 0.4
10 (3–19)
2.9 ± 0.2
19 (16–22)
CESM1-BGC
NSF-DOE-NCAR
1.25∘
2.4 ± 0.7
19 (11–27)
1.4 ± 0.2
19 (12–27)
0.9 ± 0.2
26 (14–34)
1.1 ± 0.5
10 (3–27)
1.0 ± 0.5
8 (3–16)
2.7 ± 0.1
15 (14–17)
CESM1-CAM5
NSF-DOE-NCAR
1.25∘
1.6 ± 0.3
7 (6–11)
1.4 ± 0.3
20 (9–29)
0.6 ± 0.2
8 (7–15)
1.3 ± 0.4
19 (6–30)
1.6 ± 0.4
26 (12–33)
3.0 ± 0.5
22 (12-26)
CESM1-1-FV2
NSF-DOE-NCAR
1.25∘
1.7 ± 0.1
10 (7–10)
2.1 ± 0.2
37 (32–37)
0.6 ± 0.1
7 (7–10)
1.3 ± 0.2
20 (11–27)
1.6 ± 0.2
27 (16–32)
3.9 ± 0.3
31 (28–32)
CMCC-CESM
CMCC
3.75∘
2.3 ± 0.7
17 (7–26)
2.4 ± 0.3
39 (38–41)
1.7 ± 0.5
39 (35–47)
1.8 ± 0.2
31 (29–37)
2.2 ± 0.2
38 (35–41)
3.3 ± 0.3
25 (23–27)
CMCC-CM
CMCC
0.75∘
2.3 ± 0.6
18 (10–25)
1.5 ± 0.3
23 (13–30)
1.0 ± 0.4
29 (8–35)
1.3 ± 0.2
21 (11–27)
1.6 ± 0.1
25 (18–28)
2.8 ± 0.2
17 (14–22)
CMCC-CMS
CMCC
1.8∘
2.0 ± 0.6
13 (6–22)
2.4 ± 0.3
38 (38–41)
1.1 ± 0.4
34 (12–37)
1.2 ± 0.2
14 (8–27)
1.5 ± 0.2
17 (14–29)
3.0 ± 0.3
21 (14–24)
CNRM-CM5
CNRM-CERFACS
1.4∘
3.8 ± 1.5
30 (19–41)
1.7 ± 0.4
28 (14–36)
0.9 ± 0.3
25 (8–34)
1.6 ± 0.4
30 (17–40)
1.7 ± 0.4
29 (16–35)
4.7 ± 0.9
38 (29–41)
CSIRO-Mk3-6-0
CSIRO-QCCCE
1.9∘
1.6 ± 0.2
9 (6–10)
0.8 ± 0.2
4 (3–7)
1.0 ± 0.3
32 (20–35)
1.8 ± 0.3
32 (27–42)
2.1 ± 0.4
37 (32–43)
2.5 ± 0.1
13 (11–13)
EC-EARTH
EC-EARTH
1.125∘
2.0 ± 0.4
12 (7–18)
–
–
0.8 ± 0.3
19 (7–33)
1.2 ± 0.3
11 (6–27)
1.5 ± 0.1
20 (16–28)
4.9 ± 0.4
39 (37–40)
FGOALS-g2
LASG-IAP
2.8∘
2.9 ± 0.4
25 (23–28)
1.2 ± 0.3
13 (5–27)
1.8 ± 0.4
42 (36–47)
1.8 ± 0.3
34 (28–42)
2.0 ± 0.3
34 (30–40)
3.0 ± 0.2
20 (17–23)
FIO-ESM
FIO
2.875∘
3.1 ± 0.3
27 (24-28)
1.3 ± 0.2
16 (11–25)
1.9 ± 0.2
46 (40–47)
1.9 ± 0.3
35 (28–42)
2.1 ± 0.3
36 (32–42)
2.5 ± 0.3
12 (11–16)
GFDL-CM3
NOAA GFDL
1.8∘
5.2 ± 1.0
41 (35–45)
1.2 ± 0.2
14 (8–20)
1.0 ± 0.2
27 (18–34)
1.3 ± 0.2
22 (11–27)
1.6 ± 0.1
22 (16–29)
4.4 ± 0.6
36 (31–39)
GFDL-ESM2G
NOAA GFDL
2.0∘
4.0 ± 0.9
32 (28–40)
1.2 ± 0.1
15 (12–18)
0.9 ± 0.2
22 (9–34)
1.6 ± 0.2
29 (26–33)
2.0 ± 0.2
35 (33–38)
5.6 ± 0.5
41 (40–41)
GFDL-ESM2M
NOAA GFDL
2.0∘
5.4 ± 1.4
42 (32–46)
1.5 ± 0.4
27 (12–33)
0.8 ± 0.4
17 (7–34)
1.9 ± 0.3
36 (28–42)
2.4 ± 0.4
43 (35–45)
7.1 ± 0.9
46 (42–46)
GISS-E2-H
NOAA GFDL
2.5∘
6.0 ± 1.3
45 (39–46)
1.9 ± 0.4
32 (21–37)
1.4 ± 0.3
36 (35–39)
3.2 ± 0.7
47 (46–47)
3.6 ± 0.7
46 (46–47)
9.2 ± 1.2
47 (47–47)
GISS-E2-H-CC
NOAA GFDL
2.5∘
4.1 ± 0.7
34 (29–39)
1.1 ± 0.4
9 (4–20)
1.0 ± 0.3
33 (13–35)
2.0 ± 0.4
40 (31–44)
2.3 ± 0.5
39 (32–45)
6.5 ± 0.9
42 (41–46)
GISS-E2-R
NOAA GFDL
2.5∘
4.2 ± 0.3
36 (31–37)
1.5 ± 0.3
22 (12–30)
1.0 ± 0.3
30 (15–34)
1.2 ± 0.2
15 (8–27)
1.3 ± 0.1
12 (10–15)
3.8 ± 0.5
29 (25–33)
GISS-E2-R-CC
NOAA GFDL
2.5∘
4.2 ± 0.1
35 (34–36)
1.4 ± 0.3
21 (12–29)
1.0 ± 0.3
31 (16–35)
1.3 ± 0.2
18 (11–27)
1.3 ± 0.2
14 (11–15)
4.1 ± 0.4
32 (28–36)
HadGEM2-AO
MOHC
1.25∘
4.6 ± 0.8
38 (30–42)
–
–
0.7 ± 0.2
11 (7–26)
1.6 ± 0.3
28 (18–35)
1.5 ± 0.2
16 (14–28)
4.4 ± 0.6
34 (29–39)
HadGEM2-CC
MOHC
1.25∘
4.7 ± 0.3
39 (37–40)
1.1 ± 0.1
11 (6–13)
0.8 ± 0.2
18 (8–29)
1.4 ± 0.1
27 (19–27)
1.5 ± 0.1
21 (16–28)
4.4 ± 0.3
35 (33–38)
HadGEM2-ES
MOHC
1.25∘
4.1 ± 0.7
33 (29–39)
1.1 ± 0.2
10 (5–15)
0.7 ± 0.3
12 (7–30)
1.2 ± 0.2
16 (8–27)
1.3 ± 0.2
13 (10–15)
3.8 ± 0.5
30 (27–33)
INM-CM4
INM
1.5∘
5.8 ± 0.6
44 (42–45)
2.8 ± 0.4
42 (40–43)
0.8 ± 0.2
16 (8–29)
2.4 ± 0.2
45 (43–45)
2.0 ± 0.1
33 (33–37)
4.6 ± 0.4
37 (33–39)
IPSL-CM5A-LR
IPSL
1.9∘
1.6 ± 0.6
8 (6–15)
1.5 ± 0.4
24 (9–34)
2.0 ± 0.4
47 (39–47)
2.8 ± 0.4
46 (46–47)
3.6 ± 0.4
47 (46–47)
4.3 ± 0.2
33 (32–36)
IPSL-CM5A-MR
IPSL
1.3∘
2.5 ± 0.6
22 (12–26)
1.2 ± 0.4
12 (5–27)
1.6 ± 0.4
38 (35–46)
2.0 ± 0.3
41 (31–44)
2.5 ± 0.4
45 (38–45)
3.5 ± 0.4
27 (24–30)
IPSL-CM5B-LR
IPSL
1.3∘
5.8 ± 0.7
43 (41–45)
3.8 ± 1.0
45 (42–45)
1.8 ± 0.3
45 (38–47)
2.2 ± 0.4
43 (36–45)
2.3 ± 0.2
41 (37–45)
6.8 ± 1.0
44 (42–46)
MIROC-ESM
MIROC
2.8∘
2.5 ± 0.5
21 (13–25)
1.0 ± 0.4
6 (4–20)
1.8 ± 0.2
44 (39–47)
1.4 ± 0.4
26 (8–34)
1.8 ± 0.4
32 (16–38)
2.9 ± 0.3
18 (14–24)
MIROC-ESM-CHEM
MIROC
2.8∘
2.3 ± 0.8
16 (6–26)
0.9 ± 0.4
5 (2–19)
1.8 ± 0.3
43 (38–47)
1.4 ± 0.5
23 (5–34)
1.8 ± 0.4
30 (16–37)
2.8 ± 0.4
16 (11–23)
MIROC5
MIROC
1.4∘
7.3 ± 0.4
47 (47–47)
2.6 ± 0.3
41 (38–42)
1.7 ± 0.3
40 (38–47)
2.0 ± 0.2
39 (35–42)
1.6 ± 0.1
23 (16–29)
5.2 ± 0.4
40 (39–40)
MPI-ESM-LR
MPI-M
1.9∘
4.8 ± 0.6
40 (37–41)
1.5 ± 0.3
25 (16–30)
0.7 ± 0.3
13 (7–29)
1.4 ± 0.2
25 (15–27)
1.6 ± 0.2
24 (16–31)
3.2 ± 0.2
24 (23–26)
MPI-ESM-MR
MPI-M
1.8∘
4.5 ± 0.3
37 (35–40)
1.7 ± 0.3
29 (20–34)
0.8 ± 0.4
20 (7–34)
1.3 ± 0.3
17 (8–27)
1.5 ± 0.3
18 (12–31)
3.1 ± 0.1
23 (22–24)
MRI-CGCM3
MRI
1.1∘
3.0 ± 0.3
26 (23–28)
3.2 ± 0.2
43 (43–43)
0.9 ± 0.3
24 (9–34)
1.8 ± 0.1
33 (31–37)
2.3 ± 0.2
40 (38–43)
6.7 ± 0.2
43 (42–44)
MRI-ESM1
MRI
1.1∘
2.8 ± 0.4
24 (19–28)
3.5 ± 0.4
44 (43–45)
0.9 ± 0.2
23 (10–34)
2.0 ± 0.2
38 (31–42)
2.5 ± 0.2
44 (40–45)
7.1 ± 0.3
45 (44–46)
NorESM1-M
NCC
1.9∘
1.5 ± 0.4
6 (6–11)
1.7 ± 0.2
30 (26–34)
0.7 ± 0.3
14 (7–30)
1.4 ± 0.3
24 (11–30)
1.6 ± 0.4
28 (12–33)
1.9 ± 0.1
7 (7–7)
NorESM1-ME
NCC
1.9∘
2.4 ± 0.5
20 (12–25)
1.5 ± 0.2
26 (17–30)
0.8 ± 0.2
21 (10–29)
2.0 ± 0.2
42 (34–43)
2.4 ± 0.3
42 (37–45)
2.5 ± 0.1
11 (11–13)
Measures
The climate prediction index (CPI) introduced by is
widely used in climatology studies for model evaluation and weighted
projections for example. It is
based on statistical theory for normally distributed variables, which maintains
that the probability that a realization r belongs to a population of mean
μ and a standard deviation σ is proportional to
exp(-(|r-μ|/σ)2/2). It is defined as follows:
CPIs=μsm-μso2xy/σsoxy2=rmses/σsoxy,
where the index s denotes the season, m and o
exponents are for model outputs and observations respectively,
μs is the time average of seasonal values for each grid point,
σso is the temporal standard deviation of
seasonal observation values for each grid point, 〈.〉xy is the
area-weighted spatial average, and rmses is the spatial
root mean square error for the season s.
Mean differences of sea-level pressure between models and
ERA-Interim over the period 1980–2010 (in hPa). CMIP5 model names are in
black, and reanalysis names are in blue. Hashes are for areas where the
difference is higher that 2 times the ERA-Interim annual sea-level pressure
standard deviation over the same period. External circle is 40∘ S,
and intermediate black circle is 60∘ S. Green rectangle is a typical
domain boundary for regional climate models over Antarctica
e.g.. ERA-Interim sea-level pressure over the
period 1980–2010 is displayed in the low-right panel (in hPa).
Model ranking according to CPI values: external circle is for rank 1
(ERA-Interim), while internal circle is for rank 47 (largest CPI). Models with
obvious similarities in code or produced by the same institution are marked
with the same colour (clusters), following . (a) Model rank
for winter meridional sea-ice extent (msie[win], blue diamonds), summer
sea surface temperature (tos[sum], red pentagons), annual sea-level
pressure (psl[ann], black squares), summer/winter precipitable water
(prw[s/w], black circles), summer/winter 850 hPa air
temperature (ta850[s/w], black stars), and summer
850 hPa air temperature (ta850[sum], red stars). Models are
ordered by the average of ranks. (b) Average of ranks for r1i1p1 (green dots), r2i1p1
(blue diamonds), and r3i1p1 (red squares) model realizations. When a
field was not available for the second or the third realizations, we used the CPI
value of the first realization for computing ranks. Green lines show variations
of the average of ranks when using CPIs plus/minus multi-decadal variabilities
for the considered model while not changing CPIs for other models.
Y axes: evolution in time (2070–2100 minus 1980–2010) of
summer/winter 850 hPa air temperature
(Δta850[s/w]), summer/winter precipitable water
(δprw[s/w]), summer sea surface temperature
(Δtos[sum]) and winter meridional sea-ice extent scaled by
ERA-Interim standard deviation of annual values (Δmsie[win]).
The Δ symbol is for absolute differences and the δ symbol for
absolute differences divided by 1980–2010 mean value. X axes: winter
msie bias (msie[win]b), Δta850[s/w] and
evolution in time of annual surface air temperature between 40∘ S
and 40∘ N (Δtas40S40N[ann]). Horizontal coloured lines
in the first column are 2 times the multi-decadal variability of
msie[win]b, and the grey band width is 2 times the 90th percentile of
msie[win]b multi-decadal variabilities. Solid black lines are regression
lines computed without considering the outlier BNU-ESM (red dot with black face
colour). Blue lines are vertical shift of the regression line by 1.96 standard
deviation of residuals. Three of the five highest-scores models are highlighted
with black contours: ACCESS1-3 (star), CESM1-CAM5 (thin diamond), and NorESM1-M
(triangle). Models with obvious similarities in code or produced by the same
institution are marked with the same colour, following .
When aggregating several seasons, we compute the CPI as the root mean square of
the seasonal indexes:
CPI=∑sCPIs2.
Variable selection
Our variable selection is based on three criteria: (i) the variable should be
a forcing field for RCMs, (ii) the variable should have an impact on
RCM-modelled SMB, and (iii) the variable should be constrained with sufficient
observations so that reanalyses could confidently be considered an
“observation”. Consequently, we focus on the variables detailed below.
Sea-level pressure
Sea-level pressure (psl) is a proxy for the large-scale circulation patterns
which significantly impact the precipitation patterns simulated by RCMs. The
psl spatial anomalies compared to ERA-Interim for the period 1980–2010 are
shown in Fig. . We observe that the four
seasonal psl CPIs are similar (see Fig. S1 in the Supplement), suggesting that the most relevant
metric for psl is the combination of the four seasons' CPI values, denoted by
psl[ann].
Air temperature at 850 hPa
The air temperature in the free atmosphere (here at 850 hPa; ta850) has
an impact on phase changes in RCMs (refreeze/melt of snowpack, snow/rainfall).
It also controls the maximal water vapour content of the atmosphere. Because of
its pronounced seasonal cycle, ta850 presents large temporal variability
in autumn and spring, such that seasonal means are not reliable for these
seasons, though it is more stable in summer and winter. As summer and winter
CPIs are both relevant and similar (see Fig. S1), the combined CPIs of these two
seasons form a robust metric. However, special attention should be given to
summer ta850, since it has the highest impact on the melt/refreezing
amounts and on the hydrometeors' phase changes. In conclusion, the most
relevant metrics for our study are the summer/winter ta850 CPI, denoted
by ta850[s/w], and the summer ta850 CPI, denoted by
ta850[sum].
Precipitable water
Column-integrated atmospheric water vapour, or precipitable water (prw), is
a proxy for the humidity content of the atmosphere, which impacts the amount of
precipitation in RCMs. It is affected by the same strong seasonal cycle as
temperature since the maximum water vapour content of an air parcel is related
to the temperature through the Clausius–Clapeyron relationship. Consequently,
as with ta850, seasonal prw is relevant when its value reaches its
minima and maxima, i.e. in winter and summer. Consequently we chose to focus on
the summer/winter prw CPI, which we denote by prw[s/w].
Surface oceanic conditions
Since most RCMs are not coupled with an oceanic model, sea surface temperature
(tos) and sea-ice concentration from the forcing GCM are used to
simulate oceanic conditions in the RCM's integration domain. Instead of sea-ice
concentration, we considered the meridional sea-ice extent (msie),
defined as sea-ice concentration times cell area summed for each longitude (see
Appendix B regarding normality issues). Sea-ice and open-water extents are
complementary and show very strong seasonal cycles. Consequently, seasonal
analyses for these oceanic variables should refer to winter msie CPI
(msie[win]) and summer tos CPI (tos[sum]).
Results
Multi-variable analysis
The CPI values range from 0 to ∼ 7 for msie[win] and tos[sum]
and from 0 to ∼ 3 for the other variables (Table ). In
order to obtain a global metric which gives an equal weight to each of the
variables, we first ranked the models by CPI values for each variable and then
computed the average of ranks. More oriented comparisons can be carried out
by assigning different weights to the variables of greatest interest. A
variable-by-variable comparison remains the most objective when a unique skill
score is used to evaluate a model. In Fig. a we show for
each model the ranks of its variables, with models ordered according to the
average of ranks. We evaluate the effect of multi-decadal variability of the
variables on the ranking by computing for each model and each variable the
modified rank when using CPIs plus/minus multi-decadal variabilities while not
changing CPIs for other models. Ranks and their associated ranges are detailed
in Table , and the impact on the average of ranks is displayed
in Fig. b (green lines). In addition, the average of
ranks for the first realization (r1i1p1) is similar to that of the second and third
realizations when available (Fig. b, markers), which is a
good indicator of the robustness of the method.
As expected, the five reanalyses march to the head of the podium, although the
ACCESS models perform surprisingly, with ACCESS1-3 overtaking NCEP-DOE-v2 as
well as NOAA-20CR-v2 and with ACCESS1-0 overtaking NOAA-20CR-v2. These results
are explained by the significant positive bias in precipitable water shared by
NCEP-NCAR-v1, NCEP-DOE-v2 and NOAA-20CR-v2 compared to the other reanalyses. In
addition, NOAA-20CR-v2 presents a misspecification of sea ice, with ice
concentrations never exceeding 55 % far from the coast
, which explains its low CPI for winter meridional sea-ice
extent. With regards to the other variables, the five reanalyses do not differ
significantly from ERA-Interim over 1980–2010.
Each of the CMIP5 models shows at least one variable ranked under the median
value except ACCESS1-3. The five models with the highest average ranks are
ACCESS1-3 and ACCESS1-0, although they show a significant warm bias for summer
sea surface temperature; CESM1-BGC, although it shows incorrect circulation
pattern; and CESM1-CAM5 and NorESM1-M, although they show a moderate cold bias for
summer air temperature and a wet bias for precipitable water. Two other models
have only one strong bias compared to ERA-Interim: CCSM4, showing a significant
overestimation of winter meridional sea-ice extent, and EC-EARTH, showing
a strong warm bias for summer sea surface temperature (precipitable water was
unavailable). Detailed maps of spatial anomalies relative to ERA-Interim
similar to Fig. can be found in Figs. S2 to S7.
Climate change
showed that model skills in simulating present-day
climate conditions relate only weakly to the magnitude of predicted change for
surface temperature, except for sea-ice-covered regions in winter. We looked
for emergent constraints for our region by correlating projected changes
(2079–2100 mean minus 1980–2010 mean) in winter sea-ice extent, summer sea
surface temperature, precipitable water and 850 hPa air temperature to
biases for the 1980–2010 period. We found that variable evolutions are
significantly correlated to the bias in winter sea-ice extent (p < 0.01,
Fig. , 1st column) but are poorly correlated to biases of
other variables.
Changes in precipitable water and in summer sea surface temperature are very
strongly correlated with changes in 850 hPa air temperature
(R2 > 0.8). Changes in winter sea ice are also strongly
correlated with changes in 850 hPa air temperature
(R2 = 0.68), as well as being just as well correlated with the
winter sea-ice bias (R2 = 0.62), such that these two
variables together explain more than 80 % of the variance of the
changes in winter sea ice. This suggests that studying the changes in air
temperature and in sea ice is sufficient for understanding the changes in the
four variables studied.
We introduce midlatitude (40∘ S to 40∘ N) annual surface air
temperature change as a proxy for the global warming signal. We see that
31 % of the variance of 850 hPa air temperature is explained by
the winter sea-ice bias, and almost the same amount of variance (36 %)
is explained by global warming (Fig. , 1st row), despite
winter sea-ice bias and global warming signals being uncorrelated with each
other. Additionally, changes in sea-ice extent are not significantly correlated
with the global warming signal (Fig. , 4th row). This
means that (i) the decrease in sea-ice extent is mainly driven by its simulated
state under present-day climate and that (ii) both decreasing sea-ice extent
and increasing air temperature are influenced heavily by the local feedback
between these two variables. This section highlights the importance of
simulating current climate conditions correctly, as future projected anomalies
in climate over Antarctica will be significantly dependent on the conditions of
winter sea-ice cover over the present-day period.
Discussion and conclusions
The main goal of this work was to provide a fair overview of the strengths and
weaknesses of model outputs from the last multi-model ensemble CMIP5 as a first
and essential step toward regional modelling of the Antarctic ice-sheet surface
mass balance. This study does not give an absolute ranking of CMIP5 climate
models over Antarctica as it is deliberately driven by the choice of forcing
fields for regional models. The three main factors impacting on the ranking are
the choice of reference fields, the variables selection and the measure
computation.
We chose ERA-Interim as the reference field because it has been shown to be the
most reliable contemporary global reanalysis over Antarctica
, and we included five other reanalyses
into our study to assess our knowledge of the current state of the Antarctic
climate. Our results show that these reanalyses are not significantly different
from ERA-Interim for 850 hPa air temperature, sea surface temperature,
sea-level pressure and sea-ice concentration, except for NOAA-20CR-v2, for
which sea ice was misspecified . For precipitable water,
however, we found that NCEP-NCAR-v1, NCEP-DOE-v2 and NOAA-20CR-v2 reanalyses
from NOAA share a significant positive bias when compared to ERA-Interim. This
bias was already noted by for NCEP-DOE-v2. The same
paper shows that ERA-Interim has a constant bias of -0.6 kgm-2
compared to the Special Sensor Microwave Image (SSM/I) satellite data for the 60–50∘ S area. We
compared ERA-Interim with the most recent version of satellite microwave
radiometer brightness temperatures converted to precipitable water using the
RSS Version-7 algorithm over the 1988–2014 period . We see
a bias of only -0.25 kgm-2 for the 60–50∘ S area and
of -0.21 kgm-2 for the 60–40∘ S area, for all seasons.
This bias is much lower than those encountered between ERA-Interim and models
(see Figs. S5 and S6), leading us to believe that ERA-Interim can be
confidently used as a reference for precipitable water in this region.
The variable selection is primarily based on our experience of forcing
evaluation for regional climate modelling of the Greenland Ice Sheet SMB
, with adaptations specific to the Antarctic Ice Sheet,
for which precipitation is the major component of SMB and where melt amounts
are expected to increase significantly during the century. We sought to focus
on a limited number of variables and to avoid redundancy. We considered psl
rather than 500 hPa geopotential height because the latter can be
strongly impacted by air temperature biases at low atmospheric levels, while
the centred patterns of the two variables are strongly correlated (see
Fig. S8). Another variable that could be of importance for modelling surface
mass balance is the meridional moisture flux (mmf), calculated by integrating
specific humidity times meridional wind from the surface to the top of the
atmosphere. This depends on available precipitable water as well as large-scale
circulation, driving moisture advection into the Antarctic domain. However mmf
is dominated by time-varying synoptic-scale motions, also called transient
eddies , which are captured at the sub-daily time step.
This means that a study of meridional moisture flux requires 6H outputs for all
models, which we were not able to obtain. It would be of interest to put the
vertical integral of northward and eastward water vapour flux as a standard
output in the next CMIP.
With regard to measure computation, we focused on the widely used climate
prediction index, a measure based on statistical theory for
normally distributed variables which we verified as applicable to our data set.
In order to give the same weight to the six selected variables, we chose to
first rank CMIP5 models by variable according to their CPI and then use the
average of ranks. The use of the first three realizations showed the robustness of
the ranking, after which we also evaluated the impact of multi-decadal
variability on the ranks.
In the context of these choices, ACCESS1-3 is the CMIP5 model showing the best
performance for modelling surface mass balance with a RCM. It has a significant
warm bias for summer sea surface temperature but shows no significant biases
for the five other metrics. As shown by over Greenland, biases
in sea surface temperatures only marginally impact the SMB simulated by RCMs.
In addition, ACCESS1-3 variable evolutions are close to the multi-model
ensemble mean evolutions (Fig. ). Two other models with
high skill scores could also be of particular interest because they cover the
range of plausible variable evolutions: CESM1-CAM5 and NorESM1-M, which
project future high (low) 850 hPa air temperature increase and winter
sea-ice decrease, respectively. However both models are too cold in summer,
which may impact the melt increase projected by RCMs.
With regard to climate change estimates from CMIP5, we see no significant
change in sea-level pressure patterns for RCP85 during the 21st century (see
Fig. S9), whereas the other variables evolve significantly from the 1980s to
2100. We observe that 850 hPa air temperature change combined with the
1980–2010 winter sea-ice bias explain more than 80 % of the variance
of the change in precipitable water, summer sea surface temperature and winter
sea-ice extent, while these last two variables have null correlation
with the global warming signal. This demonstrates the importance of a robust
evaluation over the current climate, as the future projected climate anomalies
over Antarctica could be significantly dependent on a model's ability to
properly simulate present-day sea-ice extent. In addition, we believe that
a better understanding of climate change over the Antarctic region would be
achieved with a better quantification of the feedback between free-atmosphere
warming and winter sea-ice decrease.
Finally, suggested that uncertainties of climate
projections over Antarctica could be better quantified by using Atmospheric Model Intercomparison Project
(AMIP)-type projections, for which sea surface conditions are computed as anomalies of the
observed state. We believe that if sea surface conditions do not improve in the
next CMIP experiment, this method would be valuable, since AMIP experiments
show reduce biases compared to historical experiments (see Fig. S10), but a
correction should be applied on anomalies to take into account the present-day
sea-ice bias of the forcing simulation.