2.1 NorESM1-F versus NorESM1-M

2.1.4 Code updates in ocean carbon cycle component

The ocean carbon cycle model coupled to MICOM is based on HAMOCC5 that originated from the work of Maier-Reimer (1993). The HAMOCC model used here has gone through several iterations of development (Maier-Reimer et al., 2005), including its first adaptation to an isopycnic ocean model (Assmann et al., 2010; Tjiputra et al., 2010). The most recent updates of the model are documented in detail in Tjiputra et al. (2013) and Schwinger et al. (2016). In NorESM1-F, in addition to the updated physical model, the updated version of HAMOCC as described by Schwinger et al. (2016) is employed. The main differences relative to the CMIP5 version (Tjiputra et al., 2013) will be discussed briefly here. As mentioned above, the time stepping of the biogeochemical tracer fields has been made fully consistent with the leapfrog time stepping of the physical fields (see Schwinger et al., 2016 for details). Further, we have activated the advanced particulate sinking scheme based on Kriest (2002), where the sinking speed of particulate organic and inorganic materials is prognostically simulated according to the particle size distributions. The model now includes several preformed tracers, such as oxygen, alkalinity, and phosphate. These preformed tracers, which are set to the respective tracer values at surface, are used to quantify biological and physical carbon pumps in the interior ocean. For air–sea gas exchange computation, the Schmidt numbers have been updated to those of Gröger and Mikolajewicz (2011).

Figure 1Annual mean time series of global mean (from top to bottom) top-of-atmosphere (TOA) net radiation, T_2 m, SST, sea surface salinity (SSS), and AMOC strength at 26.5^∘ N in the PI control run. The grey lines are annual means, whereas the black ones are 10-year running means.

Download

2.2 Experimental design

The strategy and configurations of model experiments follow Bentsen et al. (2013). The fully coupled model was first spun up for 1000 years to get into a quasi-equilibrium state with a well-ventilated upper ocean and little climate drift. The pre-industrial (PI) spin-up used aerosol emissions and concentrations of greenhouse gases defined for the year 1850 according to CMIP5 protocols. The solar constant is 1360.9 W m⁻², and CO₂ mixing ratio is set to the pre-industrial value of 284.7 ppm. The ocean component of the model was initialized from rest, and the initial ocean temperature and salinity were from the Polar Science Center Hydrographic Climatology (PHC) 3.0, updated from Steele et al. (2001). For initialization of the ocean biogeochemical fields, we use the climatological fields from the World Ocean Atlas (Garcia et al., 2010a, b) and the Global Ocean Data Analysis Project (Key et al., 2004).

Figure 2Annual mean time series of global ocean mean (a) temperature and (b) salinity. Time evolution of vertical profiles of global mean (c) temperature and (d) salinity anomalies (relative to the 1000-year mean of the PI control integration).

Download

After the PI spin-up, the simulation was integrated for another 1000 years as the PI control experiment. In this paper, we use the PI control experiment to assess model stability. A historical run was also initialized after the PI spin-up, with observation-based changes in aerosol, greenhouse gas, volcanic forcing, solar radiation, and land use for the historical time period of 1850–2005 prescribed according to the CMIP5 protocol (Bentsen et al., 2013). This historical run is used in this paper for comparison with modern observations for both the model mean state and internal variability.

In addition, two idealized CO₂ forcing experiments were initialized after the PI spin-up. The first experiment was forced with gradual CO₂ increase of 1 % per year for 140 years until quadrupling of CO₂ relative to the PI level. The second experiment was forced with abrupt quadrupling of CO₂ and was run for 150 years. These two experiments are referred to as “gradual 4 × CO₂” and “abrupt 4 × CO₂”, respectively, and are used to assess the climate sensitivity of the model.

Figure 3Time series of sea ice area (10⁶ km²) in the Northern Hemisphere and Southern Hemisphere for September and March. Black and grey lines are from the PI control experiment, and red lines are from National Snow and Ice Data Center (NSIDC) observations (Fetterer et al., 2016). The blue line is the 30-year running mean of the simulated Southern Hemisphere September sea ice area.

Download

4.1 Ocean and sea ice state

The large-scale meridional overturning circulation (MOC) in the ocean carries heat and freshwater, and plays an important role in the climate system. Modeled PI global MOC is shown in Fig. 4a. The general structure of the global MOC is similar to NorESM1-M, with a weaker Deacon cell in the Southern Ocean (22 versus 25 Sv) and a stronger counterclockwise deep circulation in the SH (13 versus 10 Sv). However, the clockwise MOC in the NH is evidently weaker than in NorESM1-M (24 versus >30 Sv), which is mainly due to weaker AMOC in NorESM1-F (Fig. 4b). As mentioned in the introduction, the strength of AMOC (20.9 Sv) was significantly reduced compared to NorESM1-M (30.8 Sv) and reached a level close to the RAPID observations at 26.5^∘ N (∼18 Sv; data from http://www.rapid.ac.uk/rapidmoc/; last access: 14 January 2019). The upper branch of AMOC in NorESM1-F is also shallower than that in NorESM1-M. Contributing to the reduced AMOC in NorESM1-F is reduced deep convection in the Labrador Sea due to stronger upper ocean restratification by the reformulated GM and modified parameterization of ocean mixed layer restratification by submesoscale eddies.

In NorESM1-M, the strong AMOC carries excessive warm and saline Atlantic waters to the high latitudes, where they are brought to depth and returned southwards in the deep Atlantic. Therefore, NorESM1-M shows a significant warm and saline bias in the deep Atlantic (see Fig. 14 in Bentsen et al., 2013). With a weakened AMOC in NorESM1-F, the warm and saline bias pattern remains in the deep Atlantic (Fig. 5), but the magnitude of the bias is moderately reduced, indicating an improved representation of water masses in the Atlantic Ocean. Ventilation is also decreased in the deep Atlantic related to the weakened AMOC in NorESM1-F, as revealed by the distribution of ideal age (see Fig. S4).

Figure 7Simulated historical (a) SST and (b) SSS biases (model minus WOA09 data; model results are averaged between 1976 and 2005).

Download

Near the ocean surface (in the upper 200 m), the fresh bias also seen in NorESM1-M remains, but the cold bias is replaced by a warm bias now in NorESM1-F. Below the sea surface (200–1000 m), there existed a cold and fresh bias both in the South and North Atlantic in NorESM1-M, whereas in NorESM1-F, the bias pattern and magnitude remain in the South Atlantic but are much mitigated in the North Atlantic. However, a prominent warm and saline bias emerges in the region around 1000 m depth and 30^∘ N in NorESM1-F which was not seen in NorESM1-M, likely reflecting the incorrect representation of pycnocline depth in the region. In the Southern Ocean, NorESM1-F shows a fresh and cold bias in the upper 1500 m that is similar to NorESM1-M, and in the lower levels, NorESM1-F shows a reasonable representation of water masses compared to the warm and saline bias seen in NorESM1-M.

Figure 8Simulated historical sea ice thickness (shading; m) averaged over the years 1976–2005 for (a) NH March, (b) NH September, (c) SH March, and (d) SH September. The red lines show the modeled mean 15 % sea ice concentration, and the thick black lines show the same from the Hadley Centre Sea Ice and Sea Surface Temperature data set (Rayner et al., 2003) for the same period.

Download

A weaker AMOC in NorESM1-F leads to a reduced northward ocean heat transport both in the Atlantic and in the global oceans relative to NorESM1-M. Compared to estimates constrained by observations, total heat transport is slightly underestimated in both hemispheres (the bias is larger in the SH; Fig. 6a). The ocean heat transport matches observationally constrained estimates well except in the region between 0 and 20^∘ S where the southward heat transport is underestimated. The slight overestimation of ocean heat transport in the NH is compensated by the slight underestimation of atmospheric heat transport. Both components of heat transport show improvement over NorESM1-M. The underestimated ocean heat transport between 0 and 20^∘ S is attributed to a too-weak southward transport in the Pacific and Indian oceans and a strong northward transport in the Atlantic Ocean (Fig. 6b). Furthermore, a northward heat transport of ∼0.6 PW is present across the Equator (it is nearly zero in observationally constrained estimates), and the surplus comes from the ocean and is mainly due to the (still) excessive northward heat transport in the Atlantic Ocean (Fig. 6b).

Simulated historical (1976–2005 mean) SST and SSS biases are shown in Fig. 7. Global mean SST in NorESM1-F is warm biased (1.09 ^∘C) relative to WOA09, compared to a cold bias of −0.15 ^∘C in NorESM1-M. The warm bias appears in most of the regions, with the strongest bias seen in the Southern Ocean, eastern boundary upwelling regions, and western boundary current regions (Gulf Stream and Kuroshio), whereas cold biases are found in the western tropical Atlantic, subpolar Atlantic region, and parts of the Nordic Seas. The cold/warm bias in the subpolar North Atlantic is likely related to the misrepresentation of the North Atlantic Current pathway; i.e., it is extended too far eastward instead of swinging northward off Newfoundland. The presented spatial pattern of SST bias resembles the simulation of NorESM1-M (Bentsen et al., 2013) and CCSM4 (Gent et al., 2011). However, compared to NorESM1-M, it seems that the much improved representation of AMOC in NorESM1-F does not bring much overall improvement in SST bias in the North Atlantic and Nordic Seas region (the cold bias is reduced while the warm bias is increased in the North Atlantic). The “isolated” strong warm biases in the Gulf Stream and Kuroshio regions indicate that the western boundary currents extend too far north before separating from the coast as is common in ocean models of similar coarse horizontal resolution (Chassignet and Marshall, 2008).

Figure 9Spatial distribution of annual mean ocean primary production from (a) model and (b) observations. The observational estimate is based on remotely sensed chlorophyll data and the Vertically Generalized Production Model (VGPM) from Behrenfeld and Falkowski (1997). Model value is taken from the last 50 years of the PI control period. Units are in g C m⁻² yr⁻¹.

Download

Simulated global mean SSS has a negative bias of −0.51 g kg⁻¹, which is larger compared to the bias of −0.15 g kg⁻¹ in NorESM1- M. Negative SSS bias occurs in most of the world's oceans except in the North Atlantic and off the Siberian coast where strong positive bias is seen (Fig. 7b). The central Arctic is featured with negative bias, as opposed to the positive bias found in NorESM1-M. Furthermore, SSS near the Weddell Sea region features a positive bias, which is likely to be associated with frequent occurrence of polynyas therein, whereby waters with higher salinity at depth are able to come to the top.

The distribution of simulated historical (1979–2005) March and September sea ice thickness and extent for both hemispheres is shown in Fig. 8. In the NH, simulated sea ice extent generally follows the observations well but is somewhat underestimated on the Pacific side of the Arctic in both seasons. On the Atlantic side, sea ice extent is slightly less than observations in the Barents Sea and Labrador Sea in March, whereas in September sea ice extent is underestimated along the periphery of the Siberian and Alaskan coasts. In the SH, sea ice extent agrees well with observations in both seasons and is only slightly underestimated.

Figure 10Taylor diagrams summarizing the statistical performance of the climatology fields of (a) phosphate, (b) oxygen, (c) dissolved inorganic carbon, and (d) total alkalinity in the water column relative to the observations (Garcia et al., 2014a, b; Lauvset et al., 2016). All standard deviations are normalized to the respective values from the observations. Different symbols represent model–data comparison at different depths from 50 to 3000 m. Purple markers represent values from NorESM1-ME pre-industrial simulations (Tjiputra et al., 2013), whereas yellow markers depict the NorESM1-F, and red circles indicate observations.

Download

Previously, NorESM1-M simulated likely too-thick sea ice in both hemispheres (Bentsen et al., 2013). The thickness is reduced in the new NorESM1-F simulation. Modeled winter sea ice thickness in the central Arctic is 1.5–2 m, which is thinner compared to the observed climatology from submarines over the years 1975–2000 (Rothrock et al., 2008) and is comparable to satellite observations over the years 2003–2008 (Kwok et al., 2009). Given the more rapidly declining trend of sea ice thickness starting from the 1990s (Kwok and Rothrock, 2009), modeled sea ice thickness is deemed to be somewhat thinner compared to observations. In addition, the reported thick sea ice bias off the East Siberian coast in NorESM1-M, that was caused by unphysical oceanic variability in high-latitude shelf regions, is significantly improved in NorESM1-F.

In the SH, a noteworthy feature of modeled September sea ice distribution is the frequent emergence of polynyas in the Weddell Sea region. Satellite observations discovered a large Weddell Sea polynya with a size of 2–3×10⁵ km² in 1974 that persisted in the following two winters (Carsey, 1980). CMIP5 models revealed that Southern Ocean polynyas are common under pre-industrial conditions but cease under anthropogenic climate forcing due to surface freshening (de Lavergne et al., 2014). In NorESM1-F, while the occurrence of Weddell Sea polynyas seems to be stochastic and irregular in both PI control and historical runs, two events of polynyas are clearly captured during the PI spin-up (see Fig. S5), each lasting for several decades. Before the opening of the polynya, ocean heat in the model is observed to gradually accumulate from below 4 km and propagate upwards in the deep ocean; then, at a tipping point, the whole water column is destabilized and heat is brought up to the ocean surface, where it melts the sea ice and deep convection initiates. The decrease of sea ice area is dramatic, i.e., $\sim \mathrm{3}\times {\mathrm{10}}^{\mathrm{6}}$ km² in the Weddell Sea polynya region (55–70^∘ S, 30^∘ W–30^∘ E), which is about 1 order of magnitude larger than the observed one in the 1970s. The production of Antarctic Bottom Water in the Atlantic sector is enhanced during the deep convection process, with an increase in volume transport of ∼2.5 Sv at 30^∘ S in the South Atlantic. Apart from the two dramatic events described above, Weddell Sea polynyas are a persistent feature (albeit with much smaller size) in both the PI and historical simulations of NorESM1-F, in contrast to NorESM1-M, which does not show any sign of polynyas. However, analysis of the September mixed layer depth and ideal age (see Figs. S6 and S7) indicates that open ocean convection is overall reduced in the Southern Ocean in NorESM1-F compared to NorESM1-M, except in the Weddell Sea polynya region and a region in the Indian section of the Southern Ocean. Thinner sea ice in NorESM1-F relative to NorESM1-M is expected to be favorable for the occurrence of Weddell Sea polynyas. Additionally, the different vertical mixing schemes and mixed layer restratification in NorESM1-F compared to NorESM1-M are also likely to play a role in creating the polynyas and also in the overall reduction of convection in the Southern Ocean.

Figure 11Monthly interannual (a) SST, (b) standard deviation of SST anomalies, and (c) skewness of SST anomalies in the NINO3.4 region for the PI control experiments (last 500 years) of NorESM1-F and NorESM1-M. Data from HadISST (averaged between 1900 and 2005) are also plotted for comparison.

Download

4.2 Ocean carbon cycle

In the last 100 years of the PI control run, most of the biogeochemical fields in the water column are in quasi-equilibrium states (note that for the sediment tracers to reach equilibrium, a much longer spin-up integration is required). The global mean net primary production (NPP) and export production are 26.8±0.3 and 4.7±0.1 PgC yr⁻¹, respectively. These values are lower than observational estimates from remote sensing as well as relative to the previous model versions (Behrenfeld and Falkowski, 1997; Tjiputra et al., 2013). Compared to the remote sensing estimates, the NPP in the model is lower because it fails to resolve the high-productivity coastal regions and too-low productivity in the oligotrophic subtropical oceans (Fig. 9). The latter is attributed to the too-low nutrient supply from subsurface, potentially related to the warm bias and too-strong stratification (Schwinger et al., 2016). Figure 9 also shows that the model is able to simulate the high production in upwelling regions of equatorial Pacific and eastern boundary upwelling systems. Despite relatively low export production, the air–sea gas exchange of CO₂ is very close to balance with small outgassing of 0.03±0.06 PgC yr⁻¹. Over the historical period, the model simulates the expected evolution of oceanic carbon sinks that correspond to higher atmospheric CO₂ concentration. Similar to NorESM1-ME (Tjiputra et al., 2013), large increase in uptake rates is pronounced after the year 1950. The simulated increasing CO₂ uptake in the 1980s and 1990s is consistent with the observational-based estimates (see Fig. S8).

Figure 12(a) Power spectra of the NINO3.4 index for the PI control experiments of NorESM1-F and NorESM1-M, with comparison to HadISST data. Composite DJF surface temperature anomalies during El Niño years for (b) NorESM1-M and (c) NorESM1-F. The data are from last 500 years of the PI control experiments for both models.

Download

Figure 10 depicts the statistical performance of the phosphate, oxygen, dissolved inorganic carbon, and alkalinity climatology fields (averaged over the last 30 years of the PI control integration), relative to the respective observational estimates, in the form of Taylor diagrams (Taylor, 2001). Also shown are values from NorESM1-ME (Tjiputra et al., 2013). For these four parameters, it is clear that the current model performance has improved noticeably from the last version, as indicated by the lower parameter biases and higher spatial correlations. These improvements are especially pronounced in the interior ocean as seen in the phosphate and oxygen tracers (except at 3000 m depth). In the previous model version, interior biases are related to the too-strong overturning circulation (Bentsen et al., 2013) and too-strong oxygen consumption for biological remineralization. These biases are now reduced.

Figure 13(a, b) Leading empirical orthogonal function (EOF) of the winter (DJF) mean sea level pressure (SLP) anomalies over the NH (20–90^∘ N) for (a) the historical run (years 1976–2005) and (b) NCEP-2 data of the same period. (c, d) Leading EOF of the monthly mean SLP anomalies over the SH (90–20^∘ S) for (c) the historical run (years 1976–2005) and (d) NCEP-2 data of the same period. The SLP patterns are obtained by regression of anomalies on the leading principal component time series. The contour intervals in all panels are 1 hPa, with the zero line omitted.

Download

5.1 The ENSO mode

To evaluate ENSO variability, we analyze the long integrations (e.g., the last 500 years) of the PI control experiments of NorESM1-F and NorESM1-M.

Monthly SST and standard deviation of monthly SST anomalies averaged over the NINO3.4 region (bounded by 120–170^∘ W and 5^∘ S–5^∘ N) are shown in Fig. 11a, b for NorESM1-F, with comparison to HadISST observations and NorESM1-M. NorESM1-M simulates a lower NINO3.4 SST and a higher variability compared to observations all year round, whereas NorESM1-F simulations feature higher NINO3.4 SST (see also the SST bias shown in Fig. 7) and weaker variability (except in September) compared to observations. Maximum NINO3.4 variability in NorESM1-F is achieved in October rather than December as in observations and NorESM1-M. The model skewness for NorESM1-F and NorESM1-M in general shows an opposite sign to observations (except for October–December in NorESM1-M), with the former showing larger nonlinearity in ENSO (Fig. 11c).

The frequency spectrum of the normalized time series of simulated detrended monthly NINO3.4 SST anomalies (Fig. 12a) shows that NorESM1-F has a narrower peak at higher frequency compared with NorESM1-M and observations.

Figure 12b, c show the composite anomalies of DJF surface temperature during El Niño years for NorESM1-F, with a comparison with NorESM1-M. An El Niño year is defined here as a year with the NINO3.4 SST anomalies greater than 1.5σ (σ is standard deviation of NINO3.4 SST anomalies) for 3 consecutive months, with at least 1 DJF month. Compared to the temperature anomalies in NorESM1-M, NorESM1-F exhibits a much weaker band of SST anomalies in the NINO3 and 4 regions; the band also extends further westward in NorESM1-F. The U-shaped negative SST anomalies surrounding the NINO regions are similar between the two experiments, with the anomalies slightly larger in the North Pacific in NorESM1-M. In the Indian Ocean, both models show cooling in the central and eastern parts and warming in the remaining majority of the ocean during El Niño years.

5.2 Northern and Southern Annular modes

The Northern Annular Mode (NAM; also known as the Arctic Oscillation) is the leading variability mode in the NH on timescales from days to decades (Thompson and Wallace, 2000). NAM is defined here as the first empirical orthogonal function (EOF) of the NH (20–90^∘ N) winter (December–February) sea level pressure (SLP) anomalies. The NAM pattern in the NorESM1-F historical experiment is shown in Fig. 13 together with the pattern derived from the National Centers for Environmental Prediction reanalysis 2 (NCEP-2) data (Kanamitsu et al., 2002). NorESM1-F generally captures the spatial pattern of NAM but also inherits similar deficiencies from NorESM1-M; e.g., NAM is stronger over the Arctic and the simulated center is migrated too far east from around Iceland to the Kara Sea, whereas simulated center of action in the North Atlantic is shifted to the east with a less symmetrical structure; the center of action over the Pacific is too strong relative to the NCEP-2 data. The SLP variance explained by NAM in NorESM1-F (29 %) is stronger than that in NCEP-2 data (22 %) but is weaker than that in NorESM1-M (36 %).

The Southern Annular Mode (SAM, also known as the Antarctic Oscillation) dominates the middle to high latitudes of the SH climate variability. It is defined here as the first EOF of the SH (90–20^∘ S) monthly SLP anomalies. The simulated spatial pattern of SAM (Fig. 13) agrees with that derived from NCEP-2 data in terms of the amplitude of the low-pressure anomalies over Antarctica, but the gradient is larger on the Pacific side; there are also some small discrepancies in amplitude and zonal asymmetry of the high-pressure anomalies. The SLP variance explained by SAM in NorESM1-F (22 %) is close to that in NCEP-2 data (25 %).