Physical model
Temperature and salinity
As Fig. but for salinity.
Zonal mean temperature (T) and salinity (S) differences between the model
and the observation-based climatology from WOA 2009
for the Atlantic and Pacific basins are shown in Figs. and
. The general patterns of T and S deviations are similar
across the different model versions and configurations. While all three
configurations include salinity relaxation, this is not balanced in the case
of Mv1, with the result that average salinity falls by 0.2 units during the
course of the integration. The predominantly negative salinity bias for the
Mv1 configuration is visible in Fig. . The mid-latitude and
tropical regions have a too strong temperature gradient in the upper
700 m, that is, a warm bias in the upper thermocline and a cold bias
below. While the magnitude and extent of the warm bias are similar for the
three model configurations (up to ≈4 ∘C), the cold bias
is weakest for Mv1, and strongest (up to -4 ∘C) for low-resolution
configuration Lv1.2.
In the Southern Ocean south of 50∘ S, the model is generally biased
towards too cold and fresh conditions (about -1 to -2 ∘C and up
to -0.5 units on the Practical Salinity Scale) below a slightly too warm
surface layer. The layer of upwelled Atlantic deep water, which is warmer and
more saline than Southern Ocean surface waters, is not or only weakly
preserved in the model. At depths below 3000–4000 m the cold and
fresh bias extends northwards into the Atlantic basin up to the Equator
(a weak cold bias extends into the North Atlantic at depth). Above these
bottom water masses and below the thermocline the Atlantic is generally too
warm by 1 to 2 ∘C and too saline by 0.2 to 0.5 units (again,
salinity is biased low in Mv1 due to the unbalanced salinity restoring flux).
In contrast, a warm and saline bias at intermediate depths in the Pacific is
mostly confined to the Southern Hemisphere, whereas the North Pacific is
biased cold and fresh at all depths below 1000 m for all model
configurations. We note that the cold and fresh bias in the Southern Ocean
water column is specific to the stand-alone configuration of the model and is
not found in the fully coupled version of NorESM
e.g.Fig. 14.
Atlantic Meridional Overturning Circulation
Atlantic Meridional Overturning Circulation (AMOC) measured as the
maximum of the meridional stream-function at 26.5∘ N for Mv1 (blue
lines), Mv1.2 (green), and Lv1.2 (red). The left panel shows the AMOC for the
period 1762–1947 during which the model is forced with the CORE normal-year
forcing. In the right panel the years 1948–2014 are shown, and thick (thin)
lines indicate the overturning simulated with the model forced by the
NCEP-C-IAF (CORE-IAF) data set. The black diamond with error bars indicates
the observational estimate provided by .
The Atlantic Meridional Overturning Circulation (AMOC) for the ocean-ice-only
configuration of Mv1.2 has been compared to results from other models in
. The forcing protocol for this latter study was to run the
model through five cycles (1948–2007) of the CORE-IAF. The AMOC strength in
our model, measured as the maximum of the annual mean meridional
streamfunction at 26.5∘ N, varied between 12.5 and 19 Sv and
showed an increasing trend of roughly 1 Svcentury-1. Under the
spin-up with the CORE-NY forcing performed in the present work, the AMOC
shows a long transient increase in strength for about 300 years before
stabilising. The average overturning for the 30 years before switching to
interannual forcing (1918–1947) is 22.3, 24.8, and 21.5 Sv for Mv1,
Mv1.2, and Lv1.2, respectively (Fig. , left panel). Aside from
the absolute values we find similar curves of AMOC strength under the forcing
protocol applied here (Fig. , right panel) compared to the
results presented in : first a 10- to 15-year decrease by 2
to 4 Sv followed by a relatively stable phase until the early 1980s,
an increase by 4 to 7 Sv towards a maximum in the late 1990s, and
another decrease until the end of the simulation period (2007 or 2014). The
Lv1.2 configuration forced by NCEP-C-IAF is an outlier in our small model
ensemble. Compared to the other model simulations, the annual- and
decadal-scale variability of AMOC strength appears to be similar but
superimposed onto a negative trend of 8 Sv over the simulation period
(see below).
In model configuration Mv1 we find only minor differences in AMOC strength
between the simulation forced with CORE-IAF and the one forced with
NCEP-C-IAF. For Mv1.2 and Lv1.2, however, the CORE-IAF simulations show
a weaker initial decrease and a generally larger overturning than the
corresponding NCEP-C-IAF simulations. These differences can be traced back to
a peculiarity of the salinity relaxation scheme when the balancing of the
relaxation flux is activated (not available in model version 1). Since too
much salt is taken out of the surface ocean globally by the unbalanced
scheme, negative salt fluxes are reduced by a multiplicative factor when
balancing of the salinity relaxation flux is activated. Interestingly, the
global restoring salt flux imbalance is larger when the model is forced with
CORE-IAF. As a consequence of correcting this imbalance, we find a stronger
reduction of the restoring salt flux out of the surface ocean in the
simulations performed with CORE-IAF compared to those forced with NCEP-C-IAF.
In the Atlantic north of 40∘ N, a positive salinity bias is
therefore reinforced in the CORE-IAF simulations with model version 1.2,
driving an increase in the AMOC relative to the simulations forced with
NCEP-C-IAF. This effect is particularly pronounced in the Lv1.2
configuration, leading to the negative AMOC trend of 8 Sv described
above. We note that the AMOC strength has been estimated as 17.2 Sv
for the time period from April 2004 to October 2012 based on observations
from the Rapid Climate Change programme (RAPID) array diamond in
Fig. . Compared to this value, our model
overestimates the AMOC strength by about 4 to 9 Sv, except for the
special case of Lv1.2 forced with NCEP-C-IAF discussed above, where AMOC is
lower than the observational estimate by about 2 Sv for the period
from 2004 to 2012.
Mixed layer depth and CFC-11
Mean seasonal cycle of mixed layer depth over the years 1961–2008
for the (a) North Atlantic, (b) North Pacific,
(c) tropics (20∘ S to 20∘ N),
(d) southern subtropics (20 to 40∘ S),
(e) latitudes between 40 and 60∘ S, and the
(f) Southern Ocean south of 60∘ S. Shown are results for
model configurations Mv1 (dark blue), Mv1.2 (green), and Lv1.2 (red). The
light blue line is the observation-based climatology of ,
which uses a density threshold criterion of 0.03 kgm-3 to define
MLD.
Seasonal cycles of modelled average bulk mixed layer depth (MLD) compared to
an observation-based MLD climatology are shown for several
regions in Fig. . The climatology uses a threshold
criterion for density; that is, MLD is defined by the depth where density has
increased by 0.03 kgm-3 relative to its near-surface value. We
note that the depth of the bulk mixed layer in our model is calculated based
on energy gain and dissipation in the surface ocean, and that modelled and
observed quantities are therefore not directly comparable. A MLD climatology
based on a density criterion is nevertheless suitable for comparison with our
model, since the criterion measures stratification directly. We find a good
agreement of modelled and observation-based MLD in terms of phasing and
summer minimum depth, whereas during autumn and winter modelled average MLDs
are up to 40 to 100 m larger outside the tropics. The largest
differences are found in the Southern Ocean south of 60∘ S
(Fig. f) during austral winter. However, the
observation-based climatology relies on very few profiles during May–October
in this region. Further, our calculation of model average MLD excluded
ice-covered grid cells, while we have no information about how many observed
profiles taken under ice cover in the Antarctic entered the climatology. We
therefore have little confidence in the model–data comparison for the
southernmost Southern Ocean during winter. In the tropics modelled average
MLDs are roughly 20 m larger than observed year round (i.e. about 60
vs. 40 m), with a small annual cycle that corresponds well to the
data-based estimate. Generally, the three model versions and configurations
show a very similar average MLD, but the low-resolution configuration Lv1.2
has a less pronounced seasonal cycle (i.e. the winter maximum MLD is
shallower and hence closer to the observation-based estimate).
Zonal mean CFC-11 concentrations (nmolm-3) along a section through
the Atlantic/Southern Ocean (as indicated by the grey shaded area in panel a) averaged over
the years 1988 to 1998 for the model configurations (a) Mv1.2, (b) Lv1.2 forced with NECP-C-IAF,
(c) Lv1.2 forced with CORE-IAF, and (d) CFC-11 concentration taken from the GLODAP gridded data
set. Panel (e) displays the global zonal mean CFC-11 column content (µmolm-2).
MLD averaged over large regions and time periods is not necessarily a useful
indicator of upper ocean ventilation since spatially and temporarily
localised convection events can transport vertically large amounts of heat
and matter. In fact, the modelled maximum MLD is frequently larger than
800 m over extended areas of the Southern Ocean (not shown). The
zonal mean distribution of CFC-11 in model version 1.2 (Fig. )
indicates a deeper than observed mixing compared to CFC-11 profiles from the
GLODAP database. In the Atlantic sector of the Southern Ocean between 500 and
1000 m depth we find relatively high average CFC-11 concentrations of
up to 3 nmolm-3, whereas the corresponding observed CFC-11
concentration is 0.44 nmolm-3. These high concentrations are
mainly caused by large fluxes occurring during Antarctic winter north of the
ice edge. We note that forcing the model with the NCEP-C-IAF data set tends
to attenuate the deep mixing in the Southern Ocean compared to simulations
carried out with the CORE-IAF (see Fig. c). In the North
Atlantic we find frequent deep convection with maximum MLDs as deep as
1400 m in the Labrador and Irminger seas as well as in the Greenland
and Norwegian seas. The comparison with GLODAP CFC-11 data reveals a tendency
towards high CFC-11 concentrations located at too large depths also in the
Labrador and Irminger seas (GLODAP does not cover the Nordic seas). At mid
latitudes we find less CFC-11 in central and intermediate water masses, best
seen as a negative bias in the modelled zonal mean integrated CFC-11 content
(Fig. e) between 40∘ S and 40∘ N.
Primary and export production
Differences between versions 1 and 1.2 of NorESM-OC are to a large extent due
to differences in the ecosystem parameter settings. We compare model results
with satellite-derived estimates of primary production (PP), which are based
on Moderate-Resolution Imaging Spectroradiometer (MODIS) data and three
different processing algorithms, the Vertically Generalized Production Model
and its Eppley variant VGPM and Eppley-VGPM, and the
Carbon based Production Model CbPM,. We processed 10
years of these data (2003–2012; obtained from
www.science.oregonstate.edu/ocean.productivity) into three
climatologies as detailed in Appendix A. Since continental shelf regions are
only poorly resolved in the model, particularly in the lower-resolution
configuration, we focus on evaluating the large-scale open ocean
productivity. We therefore exclude data from shelf regions from the following
analysis (see Appendix A for details).
Vertically integrated primary production
(molCm-2yr-1) averaged over the years 2003–2012 for model
configurations (a) Mv1, (b) Mv1.2, and (c) Lv1.2,
and (d) the mean of three satellite-based climatologies (derived
from MODIS retrievals; see text). Panel (e) displays the zonal means
of each field presented in panels (a–d). The grey shaded area
represents the range of the zonal means of the three satellite-based
climatologies.
Figure shows the mean PP over the years 2003–2012 for the
model configurations Mv1, Mv1.2, and Lv1.2 and the mean and range of the
three satellite-based climatologies. As noted in Sect. ,
model version 1 has a very strong PP at high latitudes, a feature that is not
found in the satellite-derived estimates. Modelled values are high, in excess
of 10 molCm-2yr-1 almost everywhere south of
40∘ S and in large areas north of 40∘ N
(Fig. a and e). These large discrepancies are reduced in model
version 1.2 through the re-tuning of the ecosystem model described in
Sect. . Sensitivity tests with the same physical model but
both versions of the biogeochemistry module (not shown) indicate that changes
in the physical fields between model versions do not contribute significantly
to this result. A positive bias with respect to the MODIS-based estimates is
still found south of 40∘ S, but differences are reduced to less than
5 molCm-2yr-1. North of 40∘ N, large-scale
average PP in model version 1.2 generally compares well with the
observation-based estimates, although values are now at the lower end of the
range given by the three satellite climatologies.
Primary production and export production averaged over the years 2003 to 2012
simulated by the three model versions/configurations.
Mv1
Mv1.2
Lv1.2
unit
Primary production
43.3
35.1
29.6
PgCyr-1
POC export
8.8
7.1
6.1
PgCyr-1
CaCO3 export
0.62
0.56
0.47
PgCyr-1
CaCO3 export to POC export ratio
7.1
7.9
7.8
%
Opal export
120.5
110.8
94.8
TmolSiyr-1
At low latitudes (40∘ S to 40∘ N) outside of equatorial and
coastal upwelling regions, all three model versions and configurations show
a rather low production. Since this region represents a large area of the
world ocean, the global open ocean integrated PP is lower than the mean of
the MODIS estimates of 46.3 PgCyr-1 for all model
configurations (Table ; 43.3 for Mv1, 35.1 for Mv1.2, and
29.6 PgCyr-1 for Lv1.2). Figures e and
show that this discrepancy indeed originates at the latitudes
between 40∘ S and 40∘ N. Despite the ecosystem re-tuning
and the lower nutrient consumption in the Southern Ocean in model
version 1.2, PP is only slightly higher in Mv1.2 compared to Mv1 (and about
equal in Lv1.2). We conclude that, rather than caused by inadequate ecosystem
parameter settings, the modelled low PP at low latitudes is caused by a too
low nutrient supply from below the euphotic zone, due to too stably
stratified water masses (warm bias at the surface, cold bias in the lower
thermocline, Fig. ). The lower PP in the low resolution is
consistent with the stronger cold bias found in this configuration.
Mean values of total primary production over the years 2003–2012
simulated by the three model configurations and an estimate of PP based on
satellite data (mean of three climatologies derived from MODIS retrievals;
see text) for the regions indicated in the inset. The grey bars indicate the
range of observation-based estimates.
The mean seasonal cycle of vertically integrated PP (averaged over the years
2003–2012) is shown in Fig. for six regions. The too
strong productivity in model Mv1 is reflected in a very pronounced
productivity peak in late spring or early summer everywhere south of
40∘ S and north of 40∘ N. Model version 1.2 shows
a significantly reduced annual cycle of PP in these regions (approximately
40–80 % reduction of peak production), much better in line with the
MODIS-based PP estimates. In the North Atlantic and North Pacific the peak
production of 30 molCm-2yr-1 in the updated model comes
close to the satellite-derived value. In the southern high-latitude region
(60 to 40∘ S) the modelled PP also peaks at about
30 molCm-2yr-1, but here the corresponding MODIS
estimates are much lower (peaking at 10 molCm-2yr-1). In
late summer, autumn, and winter, modelled PP at high latitudes is lower than
the observational estimate.
Mean seasonal cycle of vertically integrated primary production over
the years 2003–2012 for the (a) North Atlantic, (b) North
Pacific, (c) tropical Indian Ocean (Indian Ocean between
20∘ S and 20∘ N), (d) southern subtropics (40 to
20∘ S), (e) latitudes between 60 and 40∘ S, and
the (f) Southern Ocean south of 60∘ S. Shown are the
results for model configurations Mv1 (dark blue), Mv1.2 (green), and Lv1.2
(red) and the mean and range of the satellite-derived seasonal cycle (light
blue and grey shaded areas).
In the tropics outside the Indian Ocean there is no significant annual cycle
in the model as well as in observations. The seasonal variation in PP caused
by the monsoon in the Indian Ocean is well captured by the model, although
there appears to be a small phase shift of about 1 month relative to the
MODIS-based PP estimates (Fig. c). The subtropical annual
cycle of PP shows little difference between model versions. In the southern
subtropics (20 to 40∘ S, Fig. d, very similar
results for the northern subtropics), we find a too strong seasonal cycle,
with production peaking in early Southern Hemisphere spring
(October/November) in the model, in line with the mean of the three
MODIS-based climatologies.
Export production of POC, CaCO3, and opal
Export production of POC averaged over the years 2003–2012 for the
model configurations (a) Mv1, (b) Mv1.2, and
(c) Lv1.2; corresponding CaCO3 to organic carbon ratio in
exported matter for (d) Mv1, (e) Mv1.2, and
(f) Lv1.2; corresponding opal to organic carbon ratio in exported
matter for (g) Mv1, (h) Mv1.2, and (i) Lv1.2.
Figure shows the mean export production of POC, as well as
the CaCO3 and opal to organic carbon ratios in exported matter. The
spatial pattern of POC export closely resembles the pattern of PP, since the
fraction of PP that is exported as POC (about 15–25 %, not shown) shows
only small spatial variations. Total annual carbon export (averaged over 2003
to 2012, Table ) is 8.8 (Mv1), 7.1 (Mv1.2), and
6.1 PgCyr-1 (Lv1.2).
Since the partitioning between opal and CaCO3 export production is
parameterised dependent on available silicate in our model, CaCO3
export dominates over opal export only in regions depleted of surface
silicate, which are the subtropical gyres in the Atlantic and the South
Pacific. Due to the increased silicate to phosphorus uptake ratio
RSi:P in model version 1.2 (clearly visible in
Fig. g to i), the CaCO3 production is maintained or
even slightly expanded into the western North Pacific despite the much lower
PP (and surface silicate consumption) at high latitudes in this model
version. We note that the simple parameterisation of opal and CaCO3
export production is qualitatively supported by opal to particulate inorganic
carbon (PIC) ratios derived from sediment traps. For example, see
their Fig. 7 show that high opal/PIC ratios are constrained to
ocean regions with high surface silica concentrations, a pattern that is
qualitatively reproduced by our model.
Total modelled opal export (between 95 and 120.5 TmolSiyr-1,
Table ) is within the uncertainty range of the estimate of
105 ± 17 TmolSiyr-1 given by . The ratio
of CaCO3 export to organic carbon export of 7.1 to 7.9 % is within
the range estimated by 6 ± 3 %.
Biogeochemical tracers
Nutrients
Surface phosphate concentration (mmolm-3) averaged over the years 1965 to 2007
for the model configurations (a) Mv1, (b) Mv1.2, (c) Lv1.2, and (d) surface PO4 concentration taken
from the World Ocean Atlas gridded data set. Panel (e) displays the zonal mean of each field presented in
panels (a–d).
Maps of surface phosphate concentrations for the three model versions and
configurations Mv1, Mv1.2, and Lv1.2 as well as the observation-based global
climatology from WOA 2009 are shown in
Fig. . The reduced high-latitude PP in model version 1.2
leads to larger surface phosphate concentrations almost everywhere north of
40∘ N and south of 40∘ S compared to Mv1. This improves the
agreement with observations (i.e. reduces the negative PO4 bias found in
Mv1) in the Southern Ocean and in the North Pacific. In contrast, a positive
PO4 bias is introduced in the North Atlantic north of 40∘ N in
model version 1.2. In the Southern Ocean, simulated PO4 surface
concentrations are closest to observations in configuration Lv1.2.
Zonal mean phosphate concentration (mmolm-3) along the grey shaded region in
the Atlantic/Southern Ocean indicated in the inset in panel (a) averaged over the years 1965 to 2007
for the model configurations (a) Mv1, (c) Mv1.2, and
(e) Lv1.2. Panels (b), (d), and (f) show the differences
with respect to the World Ocean Atlas gridded data set.
In all model versions/configurations there is too much phosphate trapped at
intermediate depth in the tropical oxygen minimum zones, while there is
a negative bias below approximately 2000 m depth as well as in the
whole water column of the Southern Ocean. This general pattern is shown for
the Atlantic in Fig. , but is also found in the other ocean
basins. Besides the broad similarities, a considerable redistribution of
phosphate at depth can be observed across the different model configurations.
Due to the lower PP in the Southern Ocean in model version 1.2, less
phosphate is trapped south of 50∘ S and more phosphate is found in
the oxygen minimum zones of the tropical oceans. A similar redistribution of
nutrients in the world oceans in response to a changed export production in
the Southern Ocean has been reported earlier for other models
e.g..
As Fig. but for preformed phosphate (mmolm-3). The remineralised
fraction of phosphate is approximated by apparent oxygen utilisation (AOU) divided by the stoichiometric
oxygen to phosphorus ratio RO2:P=172 used in the model.
To investigate the different phosphate distributions further, we plot
preformed phosphate in Fig. . Here, we approximate
preformed phosphate by PO4pref≈PO4-AOU/RO2:P where RO2:P=172 is the
stoichiometric oxygen to phosphorus ratio used in the model, since
a preformed phosphate tracer was not available in version 1 of the model, and
because we can estimate preformed phosphate from WOA observations in the same
manner. The differences between the approximation and the explicitly
simulated preformed tracer (available in version 1.2) are relatively small in
the Atlantic where anaerobic remineralisation of POC is not abundant in the
model. The preformed phosphate concentrations are very similar in Mv1 and
Mv1.2 south of 50∘ S. Hence, the stronger negative bias in total
PO4 found in Mv1.2 is indeed caused by a smaller remineralised
fraction. The too strong PP in Mv1 in the Southern Ocean also results in too
low preformed phosphate concentrations in Antarctic Intermediate Waters. This
bias is clearly reduced in model version 1.2 since less nutrients are
stripped out of surface waters by biological production in the Southern
Ocean. Likewise, the increased surface PO4 concentration in model
version 1.2 in the Atlantic north of 40∘ N leads to higher than
observed preformed phosphate concentrations in intermediate and deep waters
in the Atlantic basin.
Difference [PO4]-[NO3]/16 (mmolm-3)
at the surface averaged over the years 1965 to 2007 for model configurations
(a) Mv1, (b) Mv1.2, (c) Lv1.2, and
(d) surface [PO4]-[NO3]/16 as calculated from
the World Ocean Atlas gridded data set.
The modelled distributions of nitrate and phosphate are similar in terms of
biases and general spatial structure since the model uses a fixed
stoichiometric ratio (RN : P=16) for the composition of organic
matter. The difference [PO4]-[NO3]/RN : P
(Fig. ) is positive everywhere, in broad agreement with
observations, indicating that nitrate is depleted relative to phosphate with
respect to the canonical N : P ratio of 16. Large values of this difference
are found in the tropical Pacific, in the model as well as in observations,
but are more pronounced in the model. This pattern is due to the tropical
Pacific oxygen minimum zone (OMZ) where NO3 is consumed during
denitrification to remineralise organic matter and release PO4. The
oxygen minimum zones are excessively large in our model, particularly in the
tropical Pacific (see Sect. ). These results show that the
simple parameterisation of nitrogen fixation, which occurs in the surface
ocean as soon as [NO3]<RN : P[PO4] in our
model, is active over the whole surface ocean, which is probably unrealistic.
This simple parameterisation should be viewed more as a means to keep the
model ocean close to the assumed stoichiometric N : P ratio than a
realistic parameterisation of nitrogen fixation. At depth, we find major
deviations from the similarity of nitrate and phosphate distributions in the
OMZ of the tropical Pacific. Here, our model shows a local minimum of nitrate
(due to the too strong denitrification caused by too low oxygen values)
instead of a local maximum as observed and as found for phosphate.
As Fig. but for surface silicate concentrations (mmolm-3).
Surface silicate concentrations (Fig. ) in the Southern
Ocean are much lower than observed in Mv1. Larger Si concentrations
and hence a better agreement with observations are simulated with model
version 1.2 due to the reduced PP in the Southern Ocean. Again (as for
maximum PO4 concentrations) the maximum Si values are closest
to observations in Lv1.2. The lower surface silicate concentrations
equatorward of approximately 35∘ in version 1.2 compared to version 1
is maintained (despite the much lower Si consumption in the Southern
Ocean) due to the re-tuned silicate uptake (increased RSi : P).
Although the concentration differences are not large, particularly not
relative to the half-saturation constant for silicate uptake (KSi=1mmolSim-3), the effect over more than 1000 years of
integration is a clear reduction of alkalinity in model version 1.2 (see
below) due to a slightly but constantly larger fraction of calcareous shell
formation.
Generally, the global spatial surface concentration pattern of silicate is
less well reproduced than the surface pattern of PO4 in all model
configurations. The area of high silicate concentration around Antarctica is
broader (i.e. extends further north) than observed, and the observed
north–south asymmetry with much smaller values at the northern high
latitudes (north of 40∘ N) is not well reproduced either. We note
that the pattern of elevated silicate concentrations in the North Pacific and
Arctic is similar to the pattern of PO4 concentration in the model
(Fig. ). While this is in good agreement with observations
for phosphate, observed maximum Si concentrations are 50 % smaller
than modelled maximum concentrations. This might indicate that our ecosystem
model is not tuned well enough or that its structure is oversimplified with
respect to silica cycling (e.g. fixed Si : P ratio, fixed constant sinking
speed).
Taylor diagrams for phosphate (P), nitrate (N), and silicate (S) for
model configurations Mv1 (blue circles), Mv1.2 (green triangles), and Lv1.2
(red squares) at (a) surface, (b) 500 m depth, and
(c) 2000 m depth. Panel (d) shows results for the
500 m depth level, with error-prone grid points located in the
tropics (between 20∘ N and 20∘ S) omitted from the
analysis. Observations are the objectively analysed climatologies from
WOA 2009 . Model fields have been averaged over the time
period during which the bulk of observations in WOA 2009 were acquired (1965
to 2007).
A summary of model skill for surface and interior distributions of nutrients
is given in the Taylor diagrams shown in Fig. .
Surface ocean nitrate and phosphate show high correlations with observed
fields (R>0.9 except for PO4 in Mv1, where R=0.87) and a good
agreement of spatial variability. Spatial variability is lowest in Mv1 due to
the underestimation of surface PO4 and NO3 concentrations at
high latitudes. Highest skill scores are found for Lv1.2, consistent with the
fact that this model configuration comes closest to observed surface
concentrations in the Southern Ocean. Simulated surface silicate fields show
significantly lower correlations with observations (between R=0.5 for Mv1
and R=0.8 for Lv1.2). Spatial Si variability is too high for model
version 1.2 because of the too large Si surface concentration north of
40∘ N, while Mv1 reproduces the observed surface Si
variability, however, with a wrong spatial pattern.
The improved simulation of surface nutrients in model version 1.2 comes at
the cost of less model skill in the interior ocean
(Fig. b–d). Apparently, the too strong PP in the
Southern Ocean in Mv1 partly compensates for deficiencies in the circulation
field. As discussed above, a larger Southern Ocean PP increases the amount of
phosphate trapped in the water column south of 50∘ S and reduces the
amount of phosphate trapped in the tropical oxygen minimum zones in version 1
compared to version 1.2 (Fig. ). Hence, model version 1.2
shows less correlation and a larger variability of PO4 at
500 m depth. Simulated nitrate fields at this depth level are
modified by denitrification. Since the tropical oxygen minimum zones are
considerably larger than observed, particularly in the Pacific, model skill
with respect to NO3 is reduced.
The dominant impact of the oxygen minimum zones on model skill at
500 m depth can be assessed by restricting the analysis to the
extra-tropical ocean (i.e. by omitting grid points located between
20∘ S and 20∘ N, Fig. d). Then,
phosphate variability is very close to observed, and the correlation of
modelled nitrate with observations increases by almost 0.3 for all model
configurations, such that it becomes similar to that of phosphate.
At 2000 m depth skill scores for Mv1 and Mv1.2 are similar for the
three nutrients. Lv1.2 reaches better skill scores (or about equal for
silicate) than the higher-resolution configurations. Silicate has a rather
uniform distribution at this depth and is reproduced well by all
configurations, leading to relatively high skill scores for Si.
Oxygen
As Fig. but for oxygen (molm-3) along
the grey shaded region in the Pacific/Southern Ocean indicated in the inset
in panel (a).
As discussed in Sect. there is too much deep convection
and associated tracer transport to depth in the Southern Ocean in all
NorESM-OC versions and configurations. Oxygen is particularly affected, since
O2 solubility at cold temperatures is high (this is also true for
CO2, but the upwelled water masses in the Southern Ocean tend to be rich
in DIC and low in oxygen). During the long simulations (more than 1000 years)
presented here, the deep ocean fills up with oxygen-rich waters originating
from the Southern Ocean. The resulting zonal mean O2 concentrations
in the central and eastern Pacific basin and the differences with respect to
the WOA 2009 climatology are shown in Fig. .
In the Southern Ocean south of 50∘ S and at depths deeper than
500 m, we find an average positive bias of 0.05, 0.1, and
0.09 molm-3 in Mv1, Mv1.2, and Lv1.2, respectively. The much
lower bias in model configuration Mv1 is due to the larger flux of organic
matter from the surface ocean. We find average AOU values of 0.09, 0.04, and
0.05 molm-3 (average south of 50∘ S at depths deeper
than 500 m), such that preformed O2 is virtually exactly the
same for all model versions and configurations. We note that part of the deep
ocean oxygen bias is connected to low POC transport into the deep ocean when
the standard sinking scheme is used. The large positive bias below
3000 m depth is reduced by 40 % in Lv1.2 when using alternative
sinking schemes. This is further discussed in Sect. .
The modelled oxygen minimum zones (OMZs) are excessively large in all model
versions and configurations. This is particularly true for the Pacific basin
(Fig. ), and to a lesser degree for the tropical Atlantic and
tropical Indian oceans (not shown). The impact of the parameterisation of POC
sinking on the modelled OMZs is discussed in Sect. . We
finally speculate that the strong O2 biases of opposite sign in
adjacent water masses are probably more pronounced in our isopycnic model
than they would be in a z-coordinate model, since the strong O2
gradients between these water masses are not alleviated by numerical
diffusion.
Dissolved inorganic carbon and alkalinity
Surface concentration of dissolved inorganic carbon
(molm-3) averaged over the years 1988 to 1998 for the model
configurations (a) Mv1, (b) Mv1.2, and (c) Lv1.2,
and (d) corresponding surface DIC from the observation-based GLODAP
climatology. Panel (e) displays the zonal mean of each field
presented in panels (a–d).
As Fig. but for surface total alkalinity (eqm-3).
Surface maps of DIC and alkalinity for the three model configurations and the
corresponding data-based GLODAP climatologies are shown in
Figs. and . Model version 1 has
a strong positive bias in DIC and alkalinity, which is most pronounced
between approximately 40∘ S and 50∘ N. This bias is driven
by a too large production of alkalinity at these latitudes. In our model,
alkalinity is (aside from advection and diffusion) governed by biological
production, which adds 1 mole alkalinity per mole of nitrate consumed, and
calcium carbonate production, which decreases alkalinity by 2 moles per mole
of CaCO3 produced . Since the partitioning between
opal and CaCO3 export production is parameterised dependent on
available silicate, the reduction of surface silicate concentrations at low
latitudes in model version 1.2 (Sect. ) results in
increased CaCO3 production and a considerable reduction of the
alkalinity and DIC biases. Since modelled alkalinity at low latitudes is
quite sensitive to CaCO3 export see also, we take
the reduction of alkalinity biases as indirect evidence of an improved
distribution of CaCO3 export in Mv1.2 and Lv1.2.
Average surface DIC concentrations in the Southern Ocean are lowest for
Mv1.2. Again, as in the case of phosphate, the reason for this is a lower
remineralised concentration of DIC in the upwelled water masses compared to
Mv1 and a lower preformed DIC concentration compared to Lv1.2 (not shown, but
compare Figs. and ).
Oceanic pCO2, ocean carbon sink, and anthropogenic carbon
Surface partial pressure of CO2 (µatm) averaged over
the years 1995 to 2005 for model configurations (a) Mv1,
(b) Mv1.2, (c) Lv1.2, and (d) the
observation-based surface pCO2 data set of .
Panel (e) displays the zonal mean of each field presented in
panels (a–d) and additionally the zonal mean of the data product by
. The black line indicates the mean atmospheric
pCO2 over the averaging period.
Compared to the observation-based data products of and
, we find that the general pattern of surface
CO2 partial pressure in seawater (pCO2,
Fig. ) is well reproduced by the model. The pCO2
maximum at low latitudes is broader and less confined to the equatorial
region than observed. The main factors contributing to this mismatch are the
very high primary production in the narrow eastern equatorial Pacific
upwelling band (compare Fig. ), which draws down pCO2
in the model where it has its maximum in observations, and the higher than
observed values in the upwelling areas along the western coast of America and
Africa. In the southernmost Southern Ocean, modelled mean pCO2
values are about 35 (Mv1.2) and 20 ppm (Lv1.2) lower than in the
data product. This is consistent with the DIC and alkalinity
distributions found in this region (see Figs. and
). The differences between the
pCO2 climatology and a multiannual mean of the
data are only small.
We finally use the historical and natural runs to estimate the contemporary
ocean carbon sink as well as the total storage of anthropogenic carbon
(DICant) in the ocean. The ocean carbon sink is defined here as
the carbon flux into the ocean, including its natural variability corrected
for the mean and trend found in the natural simulation. This flux is
calculated for the recent decades from 1959 to 2013 with the starting date
1959 chosen such that any shock due to the change in atmospheric forcing in
1948 has vanished and does not influence the calculation of trends. Results
for the different model versions and configurations forced with NCEP-C-IAF
and CORE-IAF are shown in Fig. .
Globally integrated annual ocean carbon sink flux 1959–2014 for Mv1 (blue lines),
Mv1.2 (green), and Lv1.2 (red). Thick (thin) lines indicate carbon fluxes simulated
with the model forced by the NCEP-C-IAF (CORE-IAF) data set. The total integrated flux from
1959 to 2007 is given in the inset, where the paler shading indicates results for simulations
with CORE-IAF. The thin black line is the carbon flux estimate by assuming
a constant 0.45 Pgyr-1 contribution of riverine outgassing flux as in .
Symbols with error bars are the carbon uptake estimates by (circle),
(square), and (diamond). Note that the estimates of
and are mean uptake values over the 1990s.
Estimates of the ocean carbon sink, accumulated carbon uptake, and
anthropogenic DIC simulated by the three model versions/configurations.
Numbers in parentheses are results from the model forced with the CORE-IAF
forcing. The second to last column gives corresponding observation-based
estimates.
Year
Mv1
Mv1.2
Lv1.2
Obs
Unit
C sink
1990–1999
2.58 (2.50)
2.33 (2.46)
2.01 (2.24)
2.0±0.44a
PgCyr-1
Acc. C uptake
1959–2007
102 (99)
93.0 (95.6)
83.7 (88.4)
—
PgC
DICant
1994
141 (139)
133 (134)
122 (123)
118±19b
PgC
DICant
2010
186
175
159
155±31c
PgC
a This is the weighted (by uncertainty) mean of the
estimates of , , and . The
given uncertainty is the root mean square of the
individual error estimates.b c
The simulated ocean carbon sink is largest for model version 1, with a mean
carbon sink of 2.58 and 2.50 PgCyr-1 during 1990 to 1999 when
forced with the NCEP-C-IAF and CORE-IAF, respectively
(Table ). The accumulated uptake over 1959 to 2007 for this
model version and the two forcing data sets differs in the same proportion
(102 vs. 99 PgC). Model version 1.2 shows a lower ocean carbon sink
with a mean value of 2.33 PgCyr-1 over the 1990 to 1999 period
for the medium-resolution configuration (2.46 when forced with CORE-IAF). The
corresponding numbers for the low-resolution simulation forced with the
NCEP-C and CORE-IAF are 2.01 and 2.24 PgCyr-1, respectively.
We compare these model results to the three independent data-based estimates
(grey symbols with error bars in Fig. ) of
, 2.0±0.4 PgCyr-1 over 1990–1999,
, 1.9±0.6 PgCyr-1 over 1990–2000, and
, 2.2±0.25 PgCyr-1 for 1995, which are the
basis for the ocean sink value given in the Intergovernmental Panel on
Climate Change Fourth Assessment Report , and which are also
used as the data-based ocean sink estimate in . The simulated
ocean sink appears to be too large in Mv1, while for version 1.2 the sink
estimates are within the error bars of the data-based estimates. The closest
match is found for the low-resolution configuration forced with NCEP-C-IAF,
which has the lowest carbon uptake. The rather too high CO2 uptake is
confirmed when comparing the model results to the CO2 flux estimates of
. There is a relatively good agreement with the CO2 sink
simulated with Lv1.2 up to 1990 and towards the end of the time series
(year 2011). From the early 1990s until 2001, the data show
a marked decline followed by a steep increase in carbon uptake, which they
attribute to a saturation followed by a reinvigoration of the Southern Ocean
carbon sink. Although our model also shows a stagnation of the ocean carbon
sink during the 1990s, there is no pronounced minimum and hence a rather
large deviation from the data around the year 2000.
Also compared to results of similar ocean biogeochemical models forced with
reanalysis data see Fig. for names and
references of individual models, the carbon uptake of NorESM-OC
appears to be large. Figure shows that, after
1980, the carbon sink in Mv1 and Mv1.2 is above the range of other models,
while Lv1.2 is at the upper end of the range.
The anthropogenic dissolved inorganic carbon (DICant) stored in the
ocean is displayed in Fig. for the year 1994, and
corresponding total numbers are given in Table . Consistent
with the larger than observed carbon sink fluxes, our model also shows higher
anthropogenic carbon storage compared to the estimate of
DICant available from the GLODAP database. In the North Atlantic
modelled anthropogenic carbon is larger than observed, which is consistent
with the too strong simulated AMOC. Model version 1.2 has a stronger AMOC
when simulations are forced with the CORE-IAF (see Sect. ),
and we find a larger accumulation of DICant in the North Atlantic
for this forcing (not shown), consistent with the larger (accumulated) uptake
flux (Fig. ). In the Southern Ocean the too high
concentration of DICant is consistent with the too strong
ventilation diagnosed from the CFC-11 tracer (Sect. ).
Larger than observed DICant concentrations are also found north of
40∘ S, where DIC is subducted in Antarctic Intermediate Waters.
Compared to a recent synthesis of anthropogenic carbon storage estimates
155±31 PgC for the year 2010, modelled
DICant storage of 186 (Mv1), 175 (Mv1.2), and 159
(Lv1.2) PgC is higher but still within the uncertainty range of the
synthesis estimate.
Particle export and sedimentation
To evaluate the three different sinking schemes available in NorESM-OC1.2, we
conducted a spin-up run over 1000 years for each scheme using the
low-resolution configuration of the model. In addition we performed one
sensitivity experiment employing the standard option, but with the constant
POC sinking speed increased from 5 to 14 md-1. These two runs
are referred to as STD-slow and STD-fast, respectively, while the model runs
employing the prognostic particle aggregation scheme and the
linear increase in sinking speed are abbreviated KR02 and WLIN in the
following text. The parameter settings used for the KR02 aggregation scheme
can be found in Table , and we use
wmin=7md-1, wmax=43md-1, and
a=40/2500md-1m-1 for the WLIN run. The global average
sinking speed profiles for the four experiments and the resulting POC fluxes
(normalised to the flux at 100 m depth) are shown in
Fig. a and b, respectively.
Faster-sinking particles provide more efficient removal of
carbon and nutrients from the surface ocean. Therefore, the amount of PP
would drop considerably in STD-fast, KR02, and WLIN compared to the standard
parameterisation if no other measures are taken. Since the PP of about
30 PgCyr-1 found for Lv1.2 is at the low end of
observation-based estimates , and since we wish to have
a similar PP in the four configurations as a starting point for our
evaluation, we tune the ecosystem parameterisation for the runs STD-fast,
KR02, and WLIN towards less efficient export and more efficient recycling of
nutrients in the euphotic zone. This is accomplished by reducing both the
fraction of grazing egested (1-ϵzoo) and the assimilation
efficiency of herbivores ωzoo (see Table ), a
measure which reduces the fraction of grazing routed to detritus and
increases the fraction that is recycled directly to phosphate
seeEqs. 9 and 11. At the end of the spin-up runs (years
1001–1010), we find globally integrated PP values of 29.6, 29.0, 28.4, and
30.4 PgCyr-1 for experiments STD-slow, STD-fast, KR02, and
WLIN. As mentioned above, the model is generally not in full equilibrium
after 1000 years, but the remaining drift is small.
Figure c shows the average POC fluxes at 100 and
2000 m depth, and at the ocean bottom (where water depth is larger
than 1000 m) for the four experiments and recent observation-based
estimates of these fluxes. The carbon export fluxes (POC flux at
100 m depth) of 4.0 to 5.6 PgCyr-1 found in the model
are at the lower end of observation-based estimates of 4.0 ,
5.7 , and 11.2 PgCyr-1 . The
three runs employing the differently tuned ecosystem show a lower export than
the STD-slow configuration. Note that in the KR02 scheme, living
phytoplankton is part of the sinking aggregates (in STD and WLIN
phytoplankton does not sink), hence the relatively large export despite the
re-tuned ecosystem. The POC flux at 2000 m is significantly smaller
than the observed range in STD-slow: 0.16 vs. 0.43 and
0.66 PgCyr-1 . It falls within or slightly
above this range for the other three runs (0.45, 0.69, and
0.68 PgCyr-1 for STD-fast, KR02, and WLIN, respectively).
A similar picture emerges for the bottom POC flux: in STD-slow it is again
much smaller than the estimate of 0.03 vs.
0.5 PgCyr-1, while the other sinking schemes yield
values which are more comparable (though smaller: 0.17, 0.32, and
0.36 PgCyr-1 for STD-fast, KR02, and WLIN, respectively). We
will evaluate the POC flux at 2000 m in more detail below.
Since the absolute POC fluxes depend on PP, and estimates of PP derived from
satellite data differ widely, we plot the export ratio or export efficiency
(EE, defined as F100POC/PP, where F100POC is the
POC flux at 100 m depth) as well as the transfer efficiency (TF,
defined as F2000POC/F100POC) in
Fig. d. The global average export efficiency in all
four model runs (STD-slow 18 %, STD-fast 14 %, KR02 21 %, and
WLIN 14 %) falls within the range estimated based on observations:
10 % , 16.3 % , and 21 %
. Consistent with the results for the absolute fluxes, the
transfer efficiency obtained for STD-slow (1.7 %) is too small in
comparison to the observational estimates of 7.5 % and
19 % , while the other runs fall within this range
(STD-fast 9 %, KR02 10 %, and WLIN 16 %). These results indicate
that the flux reaching the deep ocean below 2000 m and the sediments
is likely too low in the STD-slow configuration. Regarding the three other
schemes, a conclusion is difficult to draw, since the spread in the
observation-based estimates is large.
Globally integrated annual ocean carbon sink flux 1959–2014 for Mv1
(blue line), Mv1.2 (green), and Lv1.2 (red) forced by the NCEP-C-IAF data
set. Thin grey lines indicate the ocean carbon sink fluxes calculated by the
models used in : NEMO-PlankTOM5 , NEMO-PISCES
LSCE version,, CCSM-BEC , MPIOM-HAMOCC
, NEMO-PISCES CNRM version,, CSIRO
, and MITgcm-REcoM2 .
Vertically integrated anthropogenic carbon (molm-2) for
the year 1994 for model configurations (a) Mv1, (b) Mv1.2,
and (c) Lv1.2, and (d) corresponding DICant from
the observation-based GLODAP climatology . Panel (e)
displays the zonal mean of each field presented in panels (a–d).
(a) Mean profiles of sinking speed with depth for the four
experiments STD-slow (dark blue), STD-fast (green), KR02 (red), and WLIN
(light blue); (b) as panel (a), but for mean profiles of
POC fluxes (averaged over areas with water depth > 4000 m)
normalised to the flux at 100 m depth; (c) globally
accumulated fluxes of particulate organic carbon (POC) at 100 m depth
(dark blue), at 2000 m depth (green), and the downward POC flux at
the ocean bottom for depths larger than 1000 m (brown);
(d) export efficiency (EE, dark blue) and transfer efficiency (TE,
brown). The observation-based estimates (“Obs”) in panels (c) and
(d) are derived from , , and
for POC flux at 100 m and EE, from
and for POC flux at 2000 m and TE, and from
for the bottom POC flux. The coloured bars indicate the mean
of the observation-based estimates, while the grey error bars give the
maximum and minimum estimates. Note that the EE and TE are
calculated as ratios of total PP, total export, and total flux at
2000 m, and not as the mean of gridded EE and TE values as the rest
of the values presented here. We include these data nevertheless since the
differences between the two methods of calculation are small compared to the
spread of the observation-based data.
Global zonal mean concentrations of remineralised phosphate (mmolm-3) averaged
over the last 10 years of the spin-up simulations employing the (a) STD-slow, (c) STD-fast, (e) KR02,
and (g) WLIN particle sinking parameterisations. Panels (b, d, f, h) show the differences with respect
to the World Ocean Atlas gridded data set. The remineralised fraction of phosphate is approximated by
apparent oxygen utilisation (AOU) divided by the stoichiometric oxygen to phosphorus ratio
RO2:P=172 used in the model.
Global mean profiles of remineralised phosphate (mmolm-3) averaged
over the last 10 years of the spin-up simulations employing the STD-slow (dark blue), STD-fast (green), KR02 (red),
and WLIN (light blue) particle sinking parameterisations. The remineralised fraction of phosphate is approximated by
apparent oxygen utilisation (AOU) divided by the stoichiometric oxygen to phosphorus ratio RO2:P=172
used in the model.
The global zonal mean and global mean profiles of remineralised phosphate
(PO4remin = PO4 - PO4pref, where
PO4pref is calculated based on AOU as above) shown in
Fig. and Fig. , respectively, reveal
that there is a strong negative bias in the STD-slow simulations everywhere
below 2000 m, and also at shallower depths in the Southern Ocean and
north of 40∘ N. This bias is reduced if one of the three other
sinking schemes is used, but the reduction is weaker in the deep ocean for
STD-fast. A strong positive bias of PO4remin is found in
intermediate water masses at tropical and subtropical latitudes as well as in
the tropical oxygen minimum zones for the STD-slow scheme. This bias is not
significantly reduced in the STD-fast experiment, but there is a clear
reduction found in KR02 and WLIN. Out of the three schemes with more
efficient sinking, KR02 appears to perform slightly better in terms of
PO4remin bias, while STD-fast clearly has less skill than
KR02 and WLIN. We note that improvements in simulated
PO4remin directly translate into improved oxygen fields. The
positive O2 bias discussed in Sect. is reduced in
the deep ocean if one of the more efficient sinking schemes is employed. For
example, below 3000 m depth, average O2 biases are
0.1 molm-3 for STD-slow, 0.07 for STD-fast, and 0.06 for both
KR02 and WLIN. Oxygen minimum zones are also better simulated in KR02 and
WLIN, particularly in the Atlantic, whereas the improvement in the Pacific
basin is modest. The absence of significant amounts of remineralised
phosphate in the Arctic basin in our model (Fig. ) is due
to a combination of low PP and too strong ventilation (positive O2
biases; compare Fig. ).
Scatter plots of modelled vs. observed POC fluxes at 2000 m
depth for the POC sinking parameterisations (a) STD-slow,
(b) STD-fast, (c) KR02, and (d) WLIN. Observed
data are taken from a synthesis of sediment trap POC fluxes compiled by
. Data from several stations are averaged if they are located
in the same model grid cell. Stations are discarded if the depth in the
corresponding model grid cell is shallower than 2000 m. Model–data
pairs are colour coded according to ocean basin and latitudinal range:
Atlantic north of 40∘ N (dark blue), subtropical Atlantic (blue),
tropical Atlantic (light blue), Pacific north of 40∘ N (dark green),
subtropical Pacific (green), tropical Pacific (light green), tropical Indian
Ocean (yellow), Arabian Sea (orange), Southern Ocean between 40 and
60∘ S (brown), and Southern Ocean south of 60∘ S (red).
Here, the term “subtropical” refers to latitudes between 20 and
40∘ N as well as 20 and 40∘ S, and the term “tropical”
refers to latitudes between 20∘ N and 20∘ S. Colour codes
are also shown in the inset in panel (a).
We further evaluate the sinking schemes by comparison with a comprehensive
synthesis of sediment trap data published by . These data
have been normalised to 2000 m depth and comprise 134 stations, of
which we consider 120 here. We do not use 10 stations for which the closest
model grid point has a depth of less than 2000 m, and we disregard 2
stations in the high Arctic and 2 in the Mediterranean. Further, we form the
mean value of fluxes for stations situated in the same model grid cell,
leaving us with 102 model–data pairs. The scatter diagrams presented in
Fig. confirm the too low flux of POC in STD-low. POC
fluxes at 2000 m in this model configuration are smaller than
0.05 molCm-2yr-1 for all but eight stations in the
tropical Pacific. The latter stations are located in the narrow equatorial
tongue of excessively high PP (compare Fig. ). In this region
we find the too large oxygen minimum zones at depth, where remineralisation
by oxygen ceases and is replaced by less efficient processes
(denitrification, sulfate reduction). As mentioned above, the choice of the
sinking scheme does not significantly improve the excessive oxygen minimum
zones in the Pacific. Therefore, we also find some very high POC fluxes,
inconsistent with observations in STD-fast, KR02, and WLIN at tropical
Pacific sites (Fig. b, c, and d).
STD-fast shows modelled POC fluxes at 2000 m depth that are still
clearly too small. Except for a few sites located in the tropical Pacific,
fluxes are smaller than 0.2 molCm-2yr-1, whereas observed
values range up to 0.6 molCm-2yr-1 in the Arabian Sea.
One could argue that a further increase in the constant sinking speed would
probably further improve the modelled fluxes. However, a constant sinking
speed very efficiently removes POC and associated particular nutrients from
the upper ocean. Already the 14 md-1 constant sinking speed is
faster in the upper 500 m of the water column than the sinking speed
in KR02 and WLIN (Fig. a). A significant increase in
the constant sinking speed would therefore require strong further tuning of
ecosystem parameters towards more recycling and less efficient export to keep
PP in the range of observed values. This would lead to even smaller values
for export efficiency inconsistent with observation-based estimates.
The KR02 scheme is able to reproduce the highest POC flux values, which have
been observed in the Arabian Sea (orange dots in Fig. ).
This might be unsurprising, since the KR02 parameterisation was originally
developed and tested for application in this region. However, the scheme also
simulates rather large flux values (>0.5 molCm-2yr-1)
for three nearby stations in the tropical Indian Ocean (yellow dots in
Fig. ) and one station in the subtropical Atlantic (blue
dots), where the corresponding observed values are between 0.1 and
0.3 molCm-2yr-1. At the low end of observed flux values,
the KR02 scheme is able to reproduce the lowest observed flux values well,
but there are too many low POC flux values simulated by the scheme, mainly in
the Southern Ocean south of 60∘ S (red dots), in the Atlantic north
of 40∘ N (dark blue dots), and in the tropical Atlantic (light blue
dots).
The WLIN scheme does not yield any fluxes larger than
3.2 molCm-2yr-1 outside the tropical Pacific; that is,
the highest observed POC fluxes in the Arabian Sea are not well reproduced.
Also, the scheme does not produce any fluxes smaller than
0.02 molCm-2yr-1. Hence, three of the four smallest
observed flux values are also better simulated by KR02 than WLIN. The
correlations obtained with KR02 and WLIN are 0.32 and 0.22. However, if we
omit the high-flux stations in the Arabian Sea (station numbers 54 to 58 in
), we find correlations of 0.23 and 0.28 for KR02 and WLIN,
respectively. Hence, the relatively better correlation with sediment trap
observations in KR02 compared to WLIN is solely due to better representation
of high-POC fluxes in the Arabian Sea. If we additionally disregard the
stations in the tropical Pacific (stations 103 to 111 in ),
where the large discrepancy between model and sediment trap data is due to
other model deficiencies than the POC sinking scheme, the correlations
increase to 0.44 and 0.52 for KR02 and WLIN.
We have based the evaluation of the different POC sinking schemes on indirect
(PO4remin) and direct (sediment trap) measurements. While the
indirect method has inherent inaccuracies related to the calculation of AOU
and the assumed stoichiometry of remineralisation, the direct measurement of
POC fluxes by sediment traps also comes with large systematic uncertainties
seeand references therein. It is difficult to decide
which of the two methods provides the most reliable evaluation, and we
therefore use both approaches here. We finally note that a comprehensive
sensitivity analysis or a rigorous tuning of the different POC sinking
schemes would require accelerated offline integration techniques such as
those applied by e.g. or , which we to date
do not have available for our model.