Decadal predictions of the North Atlantic CO2 uptake

As a major CO2 sink, the North Atlantic, especially its subpolar gyre region, is essential for the global carbon cycle. Decadal fluctuations of CO2 uptake in the North Atlantic subpolar gyre region are associated with the evolution of the North Atlantic Oscillation, the Atlantic meridional overturning circulation, ocean mixing and sea surface temperature anomalies. While variations in the physical state of the ocean can be predicted several years in advance by initialization of Earth system models, predictability of CO2 uptake has remained unexplored. Here we investigate the predictability of CO2 uptake variations by initialization of the MPI-ESM decadal prediction system. We find large multi-year variability in oceanic CO2 uptake and demonstrate that its potential predictive skill in the western subpolar gyre region is up to 4–7 years. The predictive skill is mainly maintained in winter and is attributed to the improved physical state of the ocean.

REF refers to SOCAT, the root-mean-square-error is shown by the distance from the REF on the X-Axis. The results from uninitialized runs, assimilation run, initialized runs from lead time of 1-5 years, and initialized runs from lead time of 6-10 years are shown in blue, red, green, and grey, respectively.

Supplementary Note 1: Reconstruction of trends in oceanic mixing strength
The CO 2 uptake in the assimilation run shows an increasing trend in the western SPG region and a decreasing trend in the eastern SPG region (Fig. 1b); this zonal dipole trend pattern can be attributed to the NAO-related western-eastern heat loss gradient. This mechanism operates as follows. Large ocean heat loss related to a positive NAO enhances the convection activity in the western SPG region (the Labrador Sea) 4 . Hence, the mixing there is enhanced, leading to more CO 2 uptake. Meanwhile, the heat loss in the eastern SPG is much smaller than in the western SPG, the dense and saline water intrusion along the isopycnal from the west to the east weakens the mixing in the eastern SPG, thereby leading to less CO 2 uptake. Here we estimate trends of ocean mixed layer depth from 1970-1995 ( Supplementary Fig. 1). The assimilation simulation produces reverse mixing trends in the western SPG and the eastern SPG, however, the uninitialized simulations show uniform negative trend of mixing strength in the whole SPG region.

Supplementary Note 2: Predictability of SST
Before investigating the predictive skill of the oceanic carbon cycle, we estimate the predictability of ocean physical field such as SST as shown in Supplementary Fig. 2. The abrupt warming in the 1990s 5 in SPG region is well captured in the MPI-ESM-LR initialized simulations, the spread among ensemble members represented by standard deviation is relatively small. However, the uninitialized simulations mainly show a long term warming trend, and the spread among ensemble members of uninitialized simulations is much larger than that of the initialized simulations.

Supplementary Note 3: Predictability of CO 2 uptake
The temporal evolution of CO 2 uptake in the western SPG from the uninitialized simulations, assimilation simulation, and initialized simulations at lead times of 1-4 years and 2-5 years are shown in Supplementary Fig. 3. The CO 2 uptake in the uninitialized simulations shows mainly an increasing trend following the increase of anthropogenic CO 2 emissions. However, both the assimilation and the initialized simulations show considerable variations on decadal scale in addition to a long-term trend; the CO 2 uptake decreases in 1960s, increases during 1970s-1980s, and it decreases again in 1990s. The decline of CO 2 uptake in the 1990s is consistent with observational results, a rapid decline of CO 2 uptake in the subpolar North Atlantic Ocean between 1990 and 2006 was documented 6 . Moreover, the variability of CO 2 uptake, which is missing in the uninitialized simulations, can be reproduced by MPI-ESM-LR with initialization several years in advance.

Supplementary Note 4: Attribution of high predictive skill of CO 2 uptake
As shown in Fig. 3, the variations of SST and CO 2 uptake in the western SPG region are connected to each other, and their predictability may be maintained by common physical mechanism. The initialization of Atlantic meridional overturning circulation (AMOC) variability contributes to the predictive skill of SST in the SPG region 7 . It is also true for this experiment, the correlations between AMOC at 26.5°N at lead time of 1 year and SST in western SPG region at lead time from 3 years onwards are higher than 0.6 ( Supplementary   Fig. 4b). The predictive skill of CO 2 uptake in the western SPG region is also assured by initialization of the AMOC. The AMOC is highly correlated with the western SPG CO 2 flux simultaneously ( Supplementary Fig. 4a), and the correlations between AMOC and western SPG CO 2 flux are higher than 0.6 starting from lead time of 1 year till lead time of 9 years ( Supplementary Fig. 4b). To date the trend of AMOC in the past is unclear due to lack of observations. The long-term decreasing trend of AMOC in the initialized simulations may be due to the negative trend of the assimilated data from ORA-S4 which also exhibits a negative trend 8 .

Supplementary Note 5: Predictive skill of ocean surface pCO 2
Here we summarize the correlations, variability ratios, and root-mean-square-errors of model simulated ocean surface pCO 2 in the western SPG region against SOCAT observations 2 in a Taylor diagram 3 ( Supplementary Fig. 5). The correlations between SOCAT and initialized simulations at a lead time of 10 years is even larger than the correlations between SOCAT and uninitialized simulations. The root-mean-square-error against SOCAT observation are generally lower in the initialized simulations than in the uninitialized simulations. It indicates that the predictive skill of oceanic carbon cycle is improved by initializing the physical fields of the earth system model. As we use monthly mean data, the improvement may be partially but not dominantly contributed by the seasonal cycle, which is in some extent interrupted by the missing measurement and the location change of the measurements.