Natural variability of Southern Ocean convection as a driver of observed climate trends

Abstract

Observed Southern Ocean surface cooling and sea-ice expansion over the past several decades are inconsistent with many historical simulations from climate models. Here we show that natural multidecadal variability involving Southern Ocean convection may have contributed strongly to the observed temperature and sea-ice trends. These observed trends are consistent with a particular phase of natural variability of the Southern Ocean as derived from climate model simulations. Ensembles of simulations are conducted starting from differing phases of this variability. The observed spatial pattern of trends is reproduced in simulations that start from an active phase of Southern Ocean convection. Simulations starting from a neutral phase do not reproduce the observed changes, similarly to the multimodel mean results of CMIP5 models. The long timescales associated with this natural variability show potential for skilful decadal prediction. © 2018, The Author(s), under exclusive licence to Springer Nature Limited.

Full text

Articles
https://doi.org/10.1038/s41558-018-0350-3
Natural variability of Southern Ocean convection
as a driver of observed climate trends
LipingZhang 1,2*, ThomasL.Delworth1,2, WilliamCooke2,3 and XiaosongYang2,3
1Atmospheric and Oceanic Science, Princeton University, Princeton, NJ, USA. 2NOAA/Geophysical Fluid Dynamics Laboratory, Princeton, NJ, USA.
3University Corporation for Atmospheric Research, Boulder, CO, USA. *e-mail: Liping.Zhang@noaa.gov
SUPPLEMENTARY INFORMATION
In the format provided by the authors and unedited.
NATURE CLIMATE CHANGE | www.nature.com/natureclimatechange

1
Multidecadal variability of Southern Ocean convection as a
potential driver of observed Southern Ocean trends
Liping Zhang1*,2, Thomas L. Delworth1,2, William Cooke2,3 and Xiaosong Yang2,3
1Atmospheric and Oceanic Science, Princeton University, Princeton, New Jersey
2NOAA/Geophysical Fluid Dynamics Laboratory, Princeton, New Jersey
3University Corporation for Atmospheric Research, Boulder, Colorado
This document contains the following supplementary material:
(i) Observed 1974-1976 Weddell Polynya and zonal mean subsurface temperature trend over
1979-2012 (Figs. S1-S2).
(ii) Mean state and/or standard deviation of Antarctic sea ice, mixed layer depth (MLD) and the
global meridional overturning circulation in SPEAR_AM2 control run (Figs. S3-S5).
(iii) Additional details of the physical processes determining the internal multidecadal oscillation
of the Southern Ocean (SO) deep convection, including a diagnostic analysis of the influence of
SO deep convection on the atmosphere, along with the evolution of sea surface temperature (SST),
surface winds and MLD during the SO internal cycle (Figs. S6-S12).
(iv) Heat budget analysis of SO SST trend in the historical simulation started from active
convection (Fig. S13).
(v) Sea ice seasonality in observation, control run, historical simulations and SLP assimilation
(Figs. S14-15).
(vi) Persistence of SO deep convection in SPEAR_AM2 control run (Fig. S16).
Corresponding author: Liping Zhang, email: Liping.Zhang@noaa.gov

2
(i) Observed 1974-1976 Weddell Polynya and zonal mean subsurface temperature trend over
1979-2012 (Figs. S1-S2).
The Weddell Polynya is the presence of a large ice-free area within the ice-covered Weddell Sea.
The 1970s Weddell Polynya first appeared during the winter of 1974, mainly located near Maud
Rise (65oS, 0o), and persisted during the two following winters (Fig. S1a). This event is
accompanied with vigorous convective mixing and a significant mode of Antarctic Bottom water
(AABW) ventilation. After 1976 no similar polynya had been observed until 2016. We also
added a plot of modeled Weddell Polynya in the historical simulation initialized at peak
convection in Fig. S1b. The Weddell Polynya occurs around 30oW in the model, which is located
a little bit west compared to observation. The sea ice concentration between 30oW-60oE is much
smaller than that in observations and east of 20oW there is less ice in the model than observed.
We note that the mean state of sea ice concentration in the model has a low bias (Fig. S3). This
means the sea ice is more vulnerable in model than in observation even though the deep
convection induced surface warming anomalies are comparable.
Given the large uncertainty of ocean subsurface observations, we used three reanalysis data sets:
(a) Ishill1, (b) EN42, and (c) an objectively analyzed data set of annual mean ocean temperature
anomalies from the National Center for Environmental Information (NCEI)3. All data display a
surface-subsurface dipole temperature trend over 1979-2012 in the Southern Ocean (SO), with a
cooling trend at the surface and a warming trend in the subsurface (Fig. S2). These phenomena
imply a slowdown of the bottom, southern limb of the meridional overturning circulation (MOC)
between the 1980s and 2000s.
Figure S1. Observed and Modelled Weddell Polynya. (a) Observed September sea ice
concentration (SIC) averaged over 1974-1976, as calculated from HadISST data. (b) Modelled
SIC averaged in the first 3 years of the historical simulation initialized at peak convection when
the Weddell Polynya appears in any months of September, October and November. Units are
fractional sea ice concentration (a value of “1” means completely sea-ice covered).
150oW
120oW
90oW
60oW
30oW
0o
30oE
60oE
90oE
120oE
150oE
180oW
a Observed 1974−1976 mean SIC
150oW
120oW
90oW
60oW
30oW
0o
30oE
60oE
90oE
120oE
150oE
180oW
b Modelled SIC during strong convection phase
0
.2
0
.2
5 0
.
3 0
.
35 0
.4
0
.4
5 0
.
5 0
.
55 0
.
6 0
.
65 0
.7
0
.7
5 0
.
8 0
.
85 0
.
9 0
.
95
1

3
Figure S2. Zonal mean (0o-360oE) ocean temperature trend. The trends are calculated using
annual mean values from Ishii data1(a), EN42 (b) and NCEI3 (c) over the period 1979-2012.
Units are K(30yr)-1.
Zonal mean temperature trend in Ishii data
a
75S 65S 55S 45S 35S 25S 15S 5S
2
000m
1500m
1000m
500m
0m
Zonal mean temperature trend in EN4 data
b
75S 65S 55S 45S 35S 25S 15S 5S
2
000m
1500m
1000m
500m
0m
cZonal mean temperature trend in NCEI
K(30yr)−1
75S 65S 55S 45S 35S 25S 15S 5S
2
000m
1500m
1000m
500m
0m
−0.40
−0.35
−0.30
−0.25
−0.20
−0.15
−0.10
−0.05
0
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40

4
(ii) Mean state and/or standard deviation of Antarctic sea ice, mixed layer depth (MLD) and
the global meridional overturning circulation in SPEAR_AM2 control run (Figs. S3-S5).
The SPEAR_AM2 model broadly captures the observed Antarctic sea ice spatial pattern, with
relatively large (small) sea ice concentration (SIC) values in the West (East) Antarctic (Fig. S3).
The maximum SIC occurs in the Weddell Sea both in the model and observation. However, the
overall SIC magnitude in SPEAR_AM2 model is smaller than that in observations due to a warm
SST bias in the SO, related to an excess of net shortwave radiation at the surface in the model
relative to observations. In addition to the open ocean deep convection over the Weddell Sea,
Figure S4a shows that AABW formation also occurs over the continental shelves, such as the Ross
Sea and the East Antarctic shelves. Figure. S4c-d displays the ocean water age and younger water
indicates enhanced deep convection. It shows these continental shelf waters are accompanied with
subsequent downslope flow (Figure. S4c-d). Deep water formation over these shelves is a
significant advancement in the SPEAR_AM2 model compared to previous versions of GFDL
models. The largest variability of deep convection occurs over the open ocean in the Weddell Sea,
with smaller variability over the Weddell and Ross shelves (Fig. S4b). The open ocean deep
convection can covary with the convection over the continental shelves in SPEAR_AM2 model
(not shown). The global meridional overturning circulation (GMOC) in density space in the
SPEAR_AM2 model includes three main meridional cells, with the shallowest subtropical cells
(STC) in the lightest layers, the Atlantic meridional overturning circulation (AMOC) in the middle
layers and the Antarctic Bottom water (AABW) cell in the bottom densest layers (Fig. S5).
Figure S3. Long term mean Antarctic sea ice concentration. Values are calculated from (a)
SPEAR_AM2 control run and (b) HadISST data. Units are fractional sea ice concentration (a
value of “1” means completely ice-covered).
150oW
120oW
90oW
60oW
30oW
0o
30oE
60oE
90oE
120oE
150oE
180oW
150oW
120oW
90oW
60oW
30oW
0o
30oE
60oE
90oE
120oE
150oE
180oW
aModel
150oW
120oW
90oW
60oW
30oW
0o
30oE
60oE
90oE
120oE
150oE
180oW
150oW
120oW
90oW
60oW
30oW
0o
30oE
60oE
90oE
120oE
150oE
180oW
b
HadISST
0 0
.
05 0
.1
0
.1
50
.2
0
.2
5 0
.
3 0
.
35 0
.4
0
.4
5 0
.
5 0
.
55 0
.
6 0
.
65 0
.7
0
.7
5 0
.
8

5
Figure S4. Mixed layer depth (MLD) and water age in SPEAR_AM2 control run. (a) Long
term mean value and (b) standard deviation (STD) of annual-mean MLD over years 801-2000.
Units are m. It should be noted that during the first 800 years of the simulation the open ocean
MLD and its variability is considerably smaller than shown here. Open ocean convection in this
region was largely absent during the first 800 years. Vertical profile of long term mean (years
801-2000) water age along 151oE (c) and 170oE (d). Units are years. The age tracer indicates the
amount of time that has passed since the water parcel was last in contact with the atmosphere.
Lower values indicate “younger” water that was more recently ventilated. Note that the low
values for age in the water adjacent to the coast at depth indicates that, in addition to open ocean
convection, deep water is also formed as dense water on the continental shelves that flows down
the shelves to depth.
160
o
W
120oW
80oW
40
o
W
0o
40
o
E
80
o
E
120oE
160
o
E
a
m
Long term mean MLD
0 50 100 150 200 250 300 350 400
160
o
W
120oW
80oW
40
o
W
0o
40
o
E
80
o
E
120oE
160
o
E
b
m
STD of MLD
0 80 160 240 320 400 480 560 640
Water age along 151oE
c
70S 65S 55S 45S 35S 25S 15S
3
000m
2000m
1000m
0m
Water age along 170oE
d
years
70S 65S 55S 45S
3000m
2000m
1000m
0m
0 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800

6
Figure S5. Global meridional overturning circulation (GMOC) in SPEAR_AM2 control
run. Shown is the long term mean value of GMOC in density space. Unit is Sv (1 Sv = 106 m3s-
1).
(iii) Additional details of the physical processes determining the internal multidecadal
oscillation of the Southern Ocean (SO) deep convection, including a diagnostic analysis of
the influence of SO deep convection on the atmosphere, along with the evolution of sea
surface temperature (SST), surface winds and MLD during the SO internal cycle (Figs. S6-
S12).
The physical mechanisms causing the multidecadal variability of SO deep convection have much
in common with similar variability in the GFDL model CM2.14-5. Overall, the periodic
strengthening and weakening of deep convection is related to subsurface advective heating and
surface freshening. Fig. S6a shows the time evolution of annual mean temperature anomalies
averaged over the Weddell Sea. The temperature distribution is much less stratified over the
entire water column during active convection periods, while the ocean is characterized by a
dipole structure during periods with little or inactive convection (Fig. 6a, c). Here we use the
Brunt-Vaisala Frequency squared (N2) to estimate the strength of ocean stratification (Fig. 6d).
The N2 is defined as the
!"# $%
&'( )&
)*'
, where
+
is the density of sea water and g is the gravitational acceleration. N2 is
a function of temperature, salinity and pressure. The water column is relatively stable (unstable)
as the N2 becomes large (small). During inactive convection periods, heat tends to accumulate at
mid-depth (e.g. year 1580-1640 in Fig. 6a), whereas the surface becomes colder due to the strong
air-sea heat flux. The mid-depth heat spreads over time, destabilizes the ocean from below (e.g.
year 1600-1630 in Fig. S6e, the negative signal tends to propagate from mid-depth to surface),
GMOC in density space
Potential density (kg/m3)
Sv
80S 60S 40S 20S EQ 20N 40N 60N 80N
1037
1036
1035
1034
1033
1032
1031
1030
1029
−32 −28 −24 −20 −16 −12 −8−4 0 4 8 12 16 20 24 28 32

7
and eventually triggers the occurrence of deep convection (e.g. year 1640-1660). We also
estimate the relative contributions of temperature and salinity to the Brunt-Vaisala Frequency
variability. Temperature contributions are calculated by using the mean salinity and total
temperature fields, while salinity contributions are calculated by using the mean temperature and
total salinity fields. The decrease of stratification N2 from year 1600 to 1630 in the subsurface
(below 300m) is dominated by the temperature contribution (Fig. S6e), whereas the salinity
contribution plays a negative role (Fig. S6f). A heat budget analysis at 1500m reveals that the
buildup of the subsurface heat mainly arises from temperature advection (Fig. S7a), particularly
horizontal temperature advection (Fig. S7b). The horizontal temperature advection is largely
associated with the horizontal Weddell Gyre. Before the weakening of deep convection, the
westward return flow in the southern branch of the Weddell Gyre effectively brings midlatitude
heat (such as from North Atlantic Deep Water) into the Weddell Sea (not shown).!On the other
hand, the weakening of convection is preceded by subsurface heat depletion and surface
freshening as well (Fig. S6a and b). The stabilizing freshwater cap gets thicker and fresher as
convection continues to weaken (Figs. 6b-f). The minimum surface salinity occurs just slightly
before the peak phase of inactive convection (Fig. S6b). The surface salinity budget further
reveals the surface freshening before the onset of inactive convection is dominated by the surface
freshwater forcing that mainly comes from sea ice melting (Fig. S7c, d). During active
convection periods, the warm surface anomalies induce sea ice melting and excessive
precipitation (Fig. S7d). These fresh water anomalies accumulate and gradually stabilize the
ocean stratification, leading to a weakening of convection. This mechanism is explained in
greater detail in Zhang et al 20174-5.
In the SPEAR_AM2 model, the Weddell Sea dominates zonal mean characteristics across the
entire Southern Ocean. The open ocean deep convection in the Weddell Sea is also accompanied
by convection along the western coast of the Ross Sea, although the strength in the Ross Sea is
smaller than that in the Weddell Sea. The co-variability of deep convection over these two seas
may be related to the resonance, which will be further examined in future studies. In addition to
deep convection, the Ross Sea experiences positive ice-coverage-ocean-heat-storage feedback in
the upper layers that was first proposed by Lecomte et al (2017)6. Fig. S8a, b shows the
correlation between the sea ice and temperature trends in the upper 3000m over the Ross Sea in
the SPEAR_AM2 model and EN4 data2. The correlations show opposite signs between the
surface and deep ocean below 1000m. This mainly reflects the deep convection change, which
also appears in the Weddell Sea. It’s interesting to see that this anti-correlated relationship also
appears in the upper 400m, which can be seen more clearly from an enhanced image (Fig. S8c,
d). The associated physical processes are as follows: if ice formation is particularly large during
one or a few years, the brine released can be transported downward to deeper layers and not
incorporated back into the mixed layer during subsequent winters. This leads to a decrease in
surface salinity, stronger stratification, a shallower mixed layer and thus reduced vertical oceanic
heat fluxes. Thus, heat is trapped at depth in response to the larger sea ice formation, leading to a
positive sea ice-ocean feedback. This positive feedback doesn’t work for other basins in
SPEAR_AM2 model, such as the Weddell Sea and the Amundsen-Bellingshausen Seas (Fig.
S8e, f). This may be one of the reasons that the SST anomalies over the Ross Sea have a large
amplitude in response to deep convection changes in the SPEAR model. As mentioned by
Lecomte et al (2017)6, this positive feedback can be self-sustained once it is triggered by an
external perturbation.

8
It’s worth noting that in different models, the subsurface heating and surface freshening rates are
quite different due to differences in model structure, model resolution, eddy parameterization
methods and so on. Moreover, the SO mean state such as the background stratification can also
determine how much heating/freshening are needed to destabilize/stabilize the water column and
in turn trigger the occurrence/cessation of convection. As proposed by Reintges et al (2017)7,
models with a strong stratification vary on long timescales and vice versa. Thus, the period of SO
deep convection can vary from decadal to centennial time scale in different models (e.g., 140-yr
in SPEAR_AM2 and ~100-yr in GFDL CM2.14-5; Reintges et al (2017)7). Note also that the
positive sea-ice-ocean feedback that appeared over the Ross Sea in the SPEAR model and
observations may not occur in other models. The CM2.1 model did not have a similar positive
sea-ice-ocean feedback in the upper SO.
Figure S6. Heat reservoir and freshwater cap. Time evolution of annual mean (a) ocean
temperature (K) and (b) salinity (PSU) anomalies averaged over the Weddell Sea (75o–58oS,
40oW–25oE). The temperature/salinity anomaly is relative to a composite of 50 years of each of
the two major convection periods (years 1530–1580 and years 1640–1690). (c) Time series of
the AABW cell anomaly (Sv) relative to long-term mean value. (d) Evolution of the static
a Temp in the Weddell Sea
Year
K
1450 1500 1550 1600 1650 1700 1750
5000m
4000m
3000m
2000m
1000m
0m
−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
0.4
b Salt in the Weddell Sea
Year
PSU
1450 1500 1550 1600 1650 1700 1750
5000m
4000m
3000m
2000m
1000m
0m
−0.16
−0.12
−0.08
−0.04
0
0.04
0.08
0.12
0.16
1450 1500 1550 1600 1650 1700 1750
−9
−6
−3
0
3
6
9
c AABW cell anomaly
Year
Sv
d N2
Year
10−6s−2
1510 1540 1570 1600 1630 1660
1000m
800m
600m
400m
200m
−1.6
−1.2
−0.8
−0.4
0
0.4
0.8
1.2
1.6
e Temperature contribution
Year
10−6s−2
1510 1540 1570 1600 1630 1660
1000m
800m
600m
400m
200m
−1.6
−1.2
−0.8
−0.4
0
0.4
0.8
1.2
1.6
f Salinity contribution
Year
10−6s−2
1510 1540 1570 1600 1630 1660
1000m
800m
600m
400m
200m
−1.6
−1.2
−0.8
−0.4
0
0.4
0.8
1.2
1.6

9
stability N2 (
$%
&'( )&
)*
, 10-6 s-2) anomaly (relative to active convection periods), and its
contribution due to (e) temperature and (f) salinity. The dashed red (blue) lines overlapped in
figure denote the peak time of active (inactive) convection. All data are 30-yr low pass filtered to
highlight low frequency variability.
Figure S7. Heat and salinity budgets over the Weddell Sea. (a) Heat budget at 1500 m (1011 W)
averaged over the Weddell Sea (75o–58oS, 40oW–25oE) over the years 1440-1780. Red, blue, and green
lines represent advection (all directions), diffusion (horizontal plus vertical diffusions), and convection (or
call vertical mixing), respectively. The black line denotes the temperature tendency term (unit is 103 W).
(b) Temperature advection terms in three directions. Magenta, cyan, and yellow lines denote zonal,
meridional, and vertical temperature advection, respectively. (c) As in (a), but for the salinity budget (105
kg s-1) at surface. Magenta line denotes the surface salt flux forcing. The black line denotes the salinity
tendency term (unit is 10-1 kg s-1). (d) Surface salt flux forcing due to sea ice melting (cyan line), and
evaporation minus precipitation (yellow line).
1450 1500 1550 1600 1650 1700 1750
−20
−15
−10
−5
0
5
10
15
20
Year
Heat budget at 1500m
a
1450 1500 1550 1600 1650 1700 1750
−20
−15
−10
−5
0
5
10
15
20
Advection Diffusion Convection T_tendency
1450 1500 1550 1600 1650 1700 1750
−20
−15
−10
−5
0
5
10
15
20
Year
Advection in three directions
b
−uTx −vTy −wTz
1450 1500 1550 1600 1650 1700 1750
−10
−8
−6
−4
−2
0
2
4
6
8
10
Year
Salinity budget at surface
c
1450 1500 1550 1600 1650 1700 1750
−10
−8
−6
−4
−2
0
2
4
6
8
10
Advection Diffusion Convection SFLUX S_tendency
1450 1500 1550 1600 1650 1700 1750
−10
−8
−6
−4
−2
0
2
4
6
8
10
Year
Surface salt fluxes
d
EmP ice

10
Figure S8. Positive sea-ice-ocean feedback over the Ross Sea. Correlation coefficients
between sea ice concentration and ocean temperature trends in the Ross Sea (55o-80oS, 150Eo-
150oW) in (a) SPEAR_AM2 model and (b) EN4 data. The trends are calculated from the 30
years following a maximum in convective activity in the SPEAR control run and over 1979-2012
in EN4 data. Similar correlations are also found in historical simulations initialized from active
convection. This means the positive ocean sea ice feedback works in both SPEAR control run
and initialized historical simulations. (c, d) are the same as (a, b) but for the picture focused over
the upper 400m of the Ross Sea. (e, f) As in (c, d) but over the Weddell Sea (75o–58oS, 320o–
385oE) and Amundsen-Bellingshausen Seas (75o–58oS, 140oW-80oW). The green stars indicate
the correlation is significant at a 90% confidence level.
−0.6 −0.4 −0.2 0 0.2 0.4
3000m
2500m
2000m
1500m
1000m
500m
0m
Correlation
ice versus temp in Ross Sea (SPEAR)
a
−0.6 −0.4 −0.2 0 0.2 0.4 0.6
3000m
2500m
2000m
1500m
1000m
500m
0m
Correlation
ice versus temp in Ross Sea (EN4)
b
−0.6 −0.4 −0.2 0 0.2 0.4
400m
300m
200m
100m
0m
Correlation
ice versus temp in Ross Sea (SPEAR)
c
−0.6 −0.4 −0.2 0 0.2 0.4
400m
300m
200m
100m
0m
Correlation
ice versus temp in Ross Sea (EN4)
d
−0.8 −0.6 −0.4 −0.2 0 0.2
400m
300m
200m
100m
0m
Correlation
ice versus temp in Weddell Sea (SPEAR)
e
−0.8 −0.6 −0.4 −0.2 0 0.2
400m
300m
200m
100m
0m
Correlation
ice versus temp in Amundsen Sea (SPEAR)
f

11
The influence of the natural variability of SO deep convection on the atmosphere is studied in the
SPEAR_AM2 control run, using lagged Maximum Covariance Analysis (MCA). We use the
yearly MLD in the SO to highlight the low frequency variability and use three-month seasonal
means for the sea level pressure (SLP) to represent its seasonality. The remote impact of ENSO is
removed in these two variables before MCA analysis. The lagged MCA isolates pairs of spatial
patterns and their associated time series by performing a singular value decomposition of the
covariance matrix between two variables. Here the SLP and MLD are expanded into K orthogonal
spatial patterns:
,-.
/
0
1
#
2
3454
/
0
1
6'''7-8
/
0 9 :
1
#
2
;<=<
/
0 9 :
1
6'''''
>
<?@ ''
>
4?@
where
:
is the time lag,
34
and
;<
are the left and right singular vectors, and
54
and
=<
are the
time series associated with the left and right singular vectors, respectively. The homogeneous maps
for the ocean and heterogeneous maps for the atmosphere, defined as the projections of
7-8
/
0 9 :
1 and
,-.
/
0
1 onto
=<
/
0 9 :
1, is shown to study the influence of the ocean onto the
atmosphere when the ocean leads. The opposite is also true when the atmosphere leads. We use
Monte Carlo approach to test the statistical significance. A smaller significance level indicates the
presence of stronger evidence against the null hypothesis. For more details regarding the MCA
method see Czaja and Frankignoul 19998 and 20029.
Figure S9 shows the time series correlation and covariance of the first MCA mode as a function
of lag. Note that positive lags primarily reflect atmospheric forcing of the MLD, while negative
lags (longer than atmospheric persistence) can reflect the ocean feedback to the atmosphere. The
correlation and covariance are very strong during austral winter, when SLP and MLD are in phase
or SLP leads (lags>=0). This is not surprising, since the atmosphere forcing on the ocean is the
strongest and most efficient during the cold season. In contrast, the correlation and covariance are
much weaker as the MLD leads (lags<0). Small but significant correlation and covariance appear
when the MLD leads the ASO SLP by 5 years. Figure S10 further shows the MLD and SLP spatial
patterns associated with the first MCA mode. A negative Southern Annular Mode (SAM) like SLP
in ASO over the SO follows by ~5 years a deepened MLD over the open ocean Weddell Sea (Fig.
S10a). A close examination finds the SST shows a broad warming over the SO after a deepening
of MLD (Fig. S11a, b). The SST warming anomalies decrease meridional temperature gradient
and thus the lower tropospheric baroclinicity in the region of maximum eddy growth (not shown).
The decreased transient eddy heat flux and transient eddy feedback eventually lead to a negative
SAM-like SLP response. It’s worth noting that the magnitude of atmospheric response to the ocean
is small, with an order of ~ 0.1 hPa. In stark contrast, the atmosphere forcing of ocean is much
stronger when the lags are equal to or larger than zero (Fig. S10c, d), which has an order of ~ 1
hPa. This weak atmospheric response to the SO deep convection shares some similarities with the
AMOC’s influence on the North Atlantic Oscillation (NAO) in the Northern Hemisphere10. It’s
worth noting that the magnitude of the atmospheric response studied here may vary with the model
resolution. In a previous lower resolution GFDL model (GFDL-ESM2Mc), the atmospheric
response associated with SO deep convection change looks much stronger than here11. This needs
to be examined further in the future.

12
Figure S9. The first maximum covariance analysis (MCA) mode between SLP and SO
MLD. (a) Correlation and (b) Covariance (hPa m) of the associated SLP and MLD time series.
The color shades indicate the covariance significance level (%). Positive (negative) lags mean
the MLD lags (leads).
Lag (Year)
Correlation
0.25
0.2
0.15
0.11
0.09
0.11
0.07
0.11
0.09
0.07
0.07
0.09
0.05 0.05
0.05
a
−14 −12 −10 −8−6−4−2 0 2 4 6 8 10
JFM
FMA
MAM
AMJ
MJJ
JJA
JAS
ASO
SON
OND
NDJ
DJF
0
5
10
Lag (Year)
Covariance
450
400
350
300
250
200
250
200
170
140
110
200
170
140
110
110
140
110 140
b
−14 −12 −10 −8−6−4−2 0 2 4 6 8 10
JFM
FMA
MAM
AMJ
MJJ
JJA
JAS
ASO
SON
OND
NDJ
DJF
0
5
10

13
Figure S10. Spatial pattern for the first MCA mode. (a) Homogeneous map of MLD (m) and
heterogeneous map of austral winter ASO SLP (hPa) when the atmosphere lags the MLD by 5
years (Lag = -5). (b) Same as (a), but for the heterogeneous map of MLD (m) and homogeneous
map of ASO SLP (hPa) when the atmosphere and MLD is in phase (Lag = 0). (c) Same as (b)
but for when the ASO SLP leads the MLD by 4 years (Lag = 4).
80oE 140oE 160oW 100oW 40oW 20oE 80oE
80oS
70oS
60oS
50oS
40oS
30oS
a
m
0.1 0.2
0.3
02
0.3
−0.2
−0.1
−0.1
MLD leads SLP by 5 years
−320 −280 −240 −200 −160 −120 −80 −40 0 40 80 120 160 200 240 280 320
80oE 140oE 160oW 100oW 40oW 20oE 80oE
80oS
70oS
60oS
50oS
40oS
30oS
b
m
12
34
5
−1−2
−1
−1
MLD leads SLP by 0 years
−48 −42 −36 −30 −24 −18 −12 −6 0 6 12 18 24 30 36 42 48
80oE 140oE 160oW 100oW 40oW 20oE 80oE
80oS
70oS
60oS
50oS
40oS
30oS
c
m
2
11
1
−1−2
−3−4
−5
MLD lags SLP by 4 years
−48 −42 −36 −30 −24 −18 −12 −6 0 6 12 18 24 30 36 42 48

14
We show in Fig. S11 the multidecadal SST cycle associated with the SO deep convection. During
active convection, the SO experiences broad warm anomalies, with maximum values over the
Weddell Sea (Fig. S11a). These positive SST anomalies correspond to zonally oriented counter-
clockwise winds, with easterly anomalies around 40oS and westerly anomalies around 70oS. As
the convection weakens, the positive SST anomalies over the Weddell and west Ross Seas
gradually decrease, while the positive SST anomalies increase in the Amundsen-Bellingshausen
Seas (Fig. S11a-c). The warm SST over the Amundsen-Bellingshausen Seas start to weaken until
lag 24yr (Fig. S11d-f). At a lag of 40yr, negative SST anomalies emerge in the Weddell and west
Ross Seas due to the appearance of negative AABW cell anomaly (Fig. S11f). The Amundsen-
Bellingshausen Seas are still characterized by positive SST anomalies at this time. The negative
SST anomalies further grow in the Weddell and west Ross Seas and finally almost extend to the
whole SO (Fig. S11g-j). Apparently, the SST anomalies over the Amundsen-Bellingshausen Seas
lag the SST anomalies over the Weddell and west Ross Seas. This lead lag relationship is seen
more clearly from the zonal mean SST regressions exhibited in Fig. S12a. Moreover, these SST
anomalies covary with the MLD evolution (Fig. S12b). The delayed weak convection over the
Amundsen-Bellingshausen Seas is due to the advection time of salinity anomaly from the Ross
Sea (Fig. S12c). Once convection starts in the Weddell and Ross Seas, the salty water in the upper
layer is advected over adjoining marginally stable water columns, and initiates convection in them.
Although the surface wind favors southward warm advection to the Amundsen-Bellingshausen
Seas (Fig. S11-d), it can’t be the dominant factor because the magnitude of the wind response here
is so small (Fig. S10a). Meanwhile, the associated positive wind stress curl over the Amundsen-
Bellingshausen Seas generates downwelling, which favors cold SST due to suppressed
communication with subsurface warm water and thus at least partly cancels the warm effect of
advection. The delayed SST response over the Amundsen-Bellingshausen Seas shown here can
also be seen in the older GFDL CM2.14 model.

15
Figure S11. Internal cycle of SST and surface wind stress in SPEAR_AM2 control run.
Lagged regression of SST (shading, K) and surface wind stress (vector, N/m2) against the
normalized AABW cell index. All data are 30-yr low pass filtered before regression. The lag is
indicated by the “L=” notation in the upper left of each panel.
60oS
30oS
0o
30oN
60oN a L=0yra L=0yr
60oS
30oS
0o
30oN
60oN b L=8yrb L=8yr
60oS
30oS
0o
30oN
60oN c L=16yrc L=16yr
60oS
30oS
0o
30oN
60oN d L=24yrd L=24yr
60oE 120oE 180oW 120oW 60oW 0o 60oE
60oS
30oS
0o
30oN
60oN e L=32yre L=32yr
f L=40yrf L=40yr
g L=48yrg L=48yr
h L=56yrh L=56yr
i L=64yri L=64yr
60oE 120oE 180oW 120oW 60oW 0o 60oE
j L=72yr
0.01N/m2
j L=72yr
K
−0.64
−0.56
−0.48
−0.4
−0.32
−0.24
−0.16
−0.08
0
0.08
0.16
0.24
0.32
0.4
0.48
0.56
0.64

16
Figure S12. Delayed response over the Amundsen-Bellingshausen Seas. Lagged regression of
zonal mean (75o-50oS) (a) SST (K) and (b) MLD (m) against the normalized AABW cell index.
(c) Meridional mean (77oS-58oS) salinity (PSU) evolutions in the upper 700m from the Ross Sea
to the Amundsen-Bellingshausen Seas. Shown are the detrended salinity anomalies relative to
the mean over years 1000-2000.
(iv) Heat budget analysis of SO SST trend in the historical simulation started from active
convection. (Fig. S13)
To examine what physical processes cause the SST trend over the SO in the historical simulations
initialized from active convection, we performed upper 50-m averaged heat budget analysis over
three key regions, the Weddell Sea, the Ross Sea and the Amundsen-Bellingshausen Seas. Note
that all heat budget terms are calculated online and thus can be totally closed at each grid point
and each time step. In brief, the overall temperature time tendency equals to temperature tendencies
due to horizontal advection, vertical advection, vertical mixing, lateral diffusion and surface
boundary forcing. Over the Weddell and Ross Seas, the cooling SST trends are dominated by the
vertical mixing terms. This is physically consistent with the weakening of SO deep convection due
to internal cycle persistence. The warming SST trend in the Amundsen-Bellingshausen Seas is
also primarily determined by the vertical mixing term. The horizontal temperature advection also
contributes positively, but the magnitude is much smaller than the vertical mixing term.
Longitude
Lag (years)
SST regressions
a
K
150E 180W 150W 120W 90W 60W 30W 0E 30E 60E
0
5
10
15
20
25
30
35
40
45
50
−0.64
−0.48
−0.32
−0.16
0
0.16
0.32
0.48
0.64
Longitude
Lag (years)
MLD regressions
b
m
150E 180W 150W 120W 90W 60W 30W 0E 30E 60E
0
5
10
15
20
25
30
35
40
45
50
−64
−48
−32
−16
0
16
32
48
64
Salinity evolution
c
Longitude
year
PSU
160E 170W 140W 110W 80W
1000
1100
1200
1300
1400
1500
1600
1700
1800
1900
2000
−0.04
−0.03
−0.02
−0.01
0
0.01
0.02
0.03
0.04

17
Figure S13. Heat budget analysis in the historical simulations started from active convection.
(a) Southern Ocean SST trend over 1979-2012. The boxes overlapped on the trend denote the
regions for heat budget analysis. Evolution of different heat budget terms averaged over the (b)
Weddell Sea, (c) Ross Sea and (d) Amundsen-Bellingshausen Seas. Blue line denotes the
temperature tendency from horizontal advection, red line denotes the tendency due to vertical
advection, green line denotes the tendency from boundary forcing (or surface net heat flux), black
line denotes the tendency resulting from vertical mixing, yellow line denotes the tendency due to
lateral (neutral) diffusion, and magenta denotes the overall time tendency. These budget terms are
shown as anomalies with respect to their respective time mean values. Units are K(30yr)-1 for SST
trend and W/m2 for heat budget terms.
160
o
W
120oW
80
o
W
40
o
W
0o
40oE
80oE
120oE
160
o
E
a
K(30yr)−1
SST trend
−1.6
−1.4
−1.2
−1.0
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1980 1985 1990 1995 2000 2005 2010
−30
−20
−10
0
10
20
30
40
Year
Weddell Sea
b
1980 1985 1990 1995 2000 2005 2010
−10
−5
0
5
10
15
Year
Ross Sea
c
1980 1985 1990 1995 2000 2005 2010
−20
−15
−10
−5
0
5
10
15
20
Year
Amundsen−Bellingshausen Seas
d
Adv_h Adv_z Boundaryf Vmix Diffu Tendency

18
(v) Sea ice seasonality in observation, control run, historical simulations and SLP
assimilation (Figs. S14-15).
Figure S14. Sea ice concentration trend in the warm (DJFMAM) and cold (JJASON)
seasons. Observed (a) warm and (b) cold season SIC trends (100%(30yr)-1) over 1979-2012 in
NSIDC data. (c, d) Same as (a, b) but for the 30-yr SIC trend in the internal cycle in the control
run composited from active convection to the following 30 years. (e, f) Same as (a, b) but for
the SIC trend in the historical simulations initialized from active convection.
160
o
W
120oW
80oW
40
o
W
0o
40
o
E
80
o
E
120oE
160
o
E
a
Observation (NSIDC)
DJFMAM SIC trend
160
o
W
120oW
80oW
40
o
W
0o
40
o
E
80
o
E
120oE
160
o
E
bJJASON SIC trend
160
o
W
120oW
80oW
40
o
W
0o
40
o
E
80
o
E
120oE
160
o
E
c
Internal cycle
DJFMAM SIC trend
160
o
W
120oW
80oW
40
o
W
0o
40
o
E
80
o
E
120oE
160
o
E
dJJASON SIC trend
160
o
W
120oW
80oW
40
o
W
0o
40
o
E
80
o
E
120oE
160
o
E
e
Initialized from active convection
DJFMAM SIC trend
160
o
W
120oW
80oW
40
o
W
0o
40
o
E
80
o
E
120oE
160
o
E
f
100%(30yr)−1
JJASON SIC trend
−0.24 −0.21 −0.18 −0.15 −0.12 −0.09 −0.06 −0.03 0 0.03 0.06 0.09 0.12 0.15 0.18 0.21 0.24

19
Figure S15. Southern Ocean responses in SLP assimilation data. (a) The AABW cell time
series and (b) cold season SIC/SLP trends over 1979-2012 in SLP assimilation data initialized
from active convection. (c, d) Same as (a, b) but initialized from neutral convection. Units are Sv
for AABW cell index and 100%(30yr)-1 for the SIC trend. The shading in (a) and (c) denotes the
ensemble spread of AABW cell index.
(vi) Persistence of SO deep convection in SPEAR_AM2 control run (Fig. S16).
1970 1980 1990 2000 2010
8
12
16
20
Year
Sv
AABW time series
Initialized from active convection
a
1970 1980 1990 2000 2010
8
12
16
20
Year
Sv
AABW time series
Initialized from neutral convection
c
160
o
W
120oW
80oW
40
o
W
0o
40
o
E
80
o
E
120oE
160
o
E
bJJASON SIC trend
3
2
1
1
2
−5
−4
−3
−2
−1
160
o
W
120oW
80oW
40
o
W
0o
40
o
E
80
o
E
120oE
160
o
E
dJJASON SIC trend
3
21
−5
−4
−3
−2
−1
100%(30yr)−1
−0.24 −0.21 −0.18 −0.15 −0.12 −0.09 −0.06 −0.03 0 0.03 0.06 0.09 0.12 0.15 0.18 0.21 0.24

20
Figure S16. Persistence of Southern Ocean deep convection in control run. Shown are the
lagged autocorrelation of the AABW cell index. Black dash line is the 95% significance level
based on the student-t test.
References:
1. Ishii, M. & Kimoto, M. Reevaluation of historical ocean heat content variations with time-
varying XBT and MBT depth bias corrections. J. Oceanogr. 65, 287-299 (2009).
2. Good, S. A., Martin, M. J. & Rayner N. A. EN4: Quality controlled ocean temperature and
salinity profiles and monthly objective analyses with uncertainty estimates. J. Geophys. Res.
Oceans. 118, 6704–6716 (2013).
3. Levitus, S., et al. World ocean heat content and thermosteric sea level change (0- 2000 m),
1955–2010. Geophys. Res. Lett. 39, L10603 (2012).
4. Zhang, L., Delworth, T. L. & Jia, L. Diagnosis of decadal predictability of Southern Ocean sea
surface temperature in the GFDL CM2.1 model. J. Clim. 30, 6309-6328 (2017).
5. Zhang, L., et al. Estimating decadal predictability for the Southern Ocean using the GFDL
CM2.1 model. J. Clim. 30, 5187-5203 (2017).
6. Lecomte, O., et al. Vertical ocean heat redistribution sustaining sea-ice concentration trends in
the Ross Sea. Nat. Comm. 8, 258 (2017).
7. Reintges, A., Martin, T., Latif, M. & Park, W. Physical controls of Southern Ocean deep-
convectionvariability in CMIP5 models and the KielClimate Model. Geophys. Res. Lett. 44,
6951–6958 (2017).
8. Czaja, A. & Frankignoul, C. Influence of the North Atlantic SST on the atmospheric
ciuculation. Geophys. Res. Lett. 26, 2969-2972 (1999).
0 2 4 6 8 10 12 14 16 18 20 22 24 26
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Lag(Year)
AABW cell persistence
Autocorrelation
95% significance level

21
9. Czaja, A. & Frankignoul, C. Observed impact of Atlantic SST anomalies on the North Atlantic
oscillation. J. Clim. 15, 606-623 (2002).
10. Gastineau, G. & Frankignoul, C. Cold-season atmospheric response to the natural variability
of the Atlantic meridional overturning circulation. Clim. Dyn. 39, 37-57 (2012).
11. Cabré, A., Marinov, I. & Gnanadesikan, A. Global atmospheric teleconnections and
multidecadal climate oscillations driven by Southern Ocean convection. J. Clim. 30, 8107– 8126
(2017).