Tropical Forcing of Barents‐Kara Sea Ice During Autumn

The causality of the link between Autumn Barents‐Kara (BK) sea ice and the winter North Atlantic Oscillation (NAO) is uncertain, given teleconnections stemming from the tropics may influence both the extra‐tropics and the Arctic. We explore the relationship between tropical rainfall and BK sea ice in autumn, by nudging the tropics to follow observed variability in otherwise free running ensemble simulations. Tropical forcing alone can skillfully reproduce a significant fraction of observed interannual NAO variability in late autumn. We also show that interannual variability in the NAO is strongly related to simulated BK sea ice. As a result, we are able to reproduce some of the observed link between tropical rainfall and autumn BK sea ice. However, only during the strong 1997 El Niño are clear tropical influences at high latitudes found. Large ensembles and strong tropical forcing are required to detect tropical forced variability in models at high latitudes.

• Tropical nudging experiments reproduce some of the observed interannual variability in autumn Barents-Kara sea ice • Tropical nudging reproduces a significant fraction of observed autumn North Atlantic Oscillation (NAO) variability, and the anticorrelation of NAO with sea ice • Only during the strong El Niño of 1997 is the model able to reproduce the strong observed teleconnection from the tropics to sea ice is supported by studies exploring multi-model ensembles such as Kelleher and Screen (2018), who show in CMIP5 simulations that the link between winter Arctic sea-ice and Eurasian temperature is only strong when the atmosphere leads the sea ice. Physically, this can be explained through wind patterns modulating sea ice export out of the Arctic basin, or through transporting heat between the atmosphere and the ocean (Guemas et al., 2016;McCusker et al., 2016). Studies by Ding et al. (2022) and Olonscheck et al. (2019) have also quantified the contribution of atmospheric variability on sea ice in a controlled way. Sea ice is also sensitive to high frequency weather variability, where storms can break up large regions of ice and impact on regional scale Arctic sea ice coverage (Guemas et al., 2016;Kohout et al., 2014). Thus, a positive NAO event may precede anomalously low Arctic sea ice through poleward advection of heat, moisture, cloud and increased ocean mixing/ice breakup along with latitudinal ice transport.
Despite studies demonstrating the large role extratropical atmospheric variability plays in Arctic sea-ice variability, late autumn Arctic sea-ice has also been used as a predictor of the winter extratropical circulation in observations. Empirical predictions have used autumn Arctic sea-ice to forecast the winter NAO (Garcia-Serrano et al., 2015;Hall et al., 2017;Wang et al., 2017), as well as Eurasian winter temperatures (Wu et al., 2011). While some studies support a physical mechanism linking autumn sea ice variability to the winter extratropical atmosphere (Honda et al., 2009;Kim et al., 2014;Mori et al., 2019), many studies find no significant link between autumn Arctic sea-ice and the winter circulation. This is typically due to large internal variability, which requires very large ensembles to isolate a signal driven by sea ice changes (Screen, 2017;Siew et al., 2021;Smith et al., 2022;R. Zhang & Screen, 2021). Additionally, sensitivities in experimental setup such as the inclusion of ocean coupling (Blackport & Kushner, 2018), model errors in simulating ocean-ice variability or differences the background state of the model Strommen et al., 2022) are likely factors contributing to the divergence in results here. An additional complexity is non-stationarity: Kolstad and Screen (2019) found considerable decadal variability in the correlation between BK sea ice and the winter NAO, which means long time series are required to identify robust relationships.
We focus our study on determining whether there is a link between tropical variability and BK sea ice during autumn. This is motivated by several factors. First, many studies show a tropical impact on sea ice variability in boreal winter, but to the authors knowledge, none have investigated this link in autumn. Second, previous work has shown that the well documented link between autumn Arctic sea ice and winter extratropical variability can be explained, at least partly, through tropical forcing (Warner et al., 2020). As such, given the persistence of tropical drivers such as ENSO which can extend through multiple seasons, it is plausible that tropical variability may impact autumn sea ice variability as well as winter extratropical variability. Interactions between the extra-tropical atmosphere and Arctic sea ice in autumn may also be important in preconditioning winter weather (King & García-Serrano, 2016), although again these may not be directly causal and may simply be a dual response to tropical variability.
We explore the potential role that the tropics has on both extratropical and high-latitude variability during autumn months, by performing tropical nudging experiments that allow us to estimate how much high-latitude interannual variability can be attributed to the tropics. This work builds on studies such as Jung et al. (2014), Hansen et al. (2017), and Maidens et al. (2020) who looked at the impact of tropical nudging on winter high-latitude skill from a numerical weather prediction perspective.

Methods
We use the Hadley Center Global Environmental Model version 3 with the Global Coupled model 2.0 configuration (HadGEM3-GC2; Williams et al., 2015). This has previously been used in the Decadal Prediction System 3 system shown by Dunstone et al. (2016) to skillfully predict both the NAO and Arctic sea ice, and is based on the GloSea5-GC2 configuration. The model has a horizontal resolution of 0.83° × 0.55° and 85 levels with a lid at 85 km. A timestep of 15 min is used. It employs interactive sea ice (CICE; Rae et al., 2015) and ocean (NEMO; Madec et al., 2017). We initialize the model using the observational conditions from the 1 September 2000 in all cases to prevent any differences in initial conditions (Uotila et al., 2019). We choose this year because autumn 2000 is a typical season when evaluated within the period 1980-2017, where pan-Arctic and BK total sea ice area is within the one standard deviation of the mean. Additionally, while tropical conditions indicate a weak La Nina, tropical rainfall in all four basins (as defined in Scaife et al. (2017)) are also within 1 standard deviation of 1980-2017 mean conditions. Model simulations are run for 3 months through to the start of December.
Model simulations are performed, nudging the tropics to the observationally-constrained reanalysis for the years 1993-2015 inclusive. For each year, a 20-member ensemble is produced using stochastic perturbations, giving 460 simulations in total. In each case, we nudge the tropical column at every timestep to ERA-Interim reanalysis (Dee et al., 2011) for the respective year. The tropical relaxation spans all vertical model levels bar the lowest model level, and extends ±19.5° north/south of the equator, with an additional 8° latitude tapering. The strength of the nudging reduces linearly with latitude over the tapering. Potential temperature, along with the zonal and meridional components of wind are nudged at every timestep, with an e-folding timescale of 6 hr (Telford et al., 2008). Nudging moisture fields has little impact, so is not included. This configuration allows us to prescribe tropical variability and infer its impact on the mid and high-latitude circulation during autumn in a controlled manner, while accounting for internal variability using the ensemble.
In order to verify the model simulations, the Hadley Center Sea Ice and Sea Surface Temperature data set version 2 (HADISST2; Titchner & Rayner, 2014) is used for monthly sea ice concentrations. For precipitation, we use the Global Precipitation Climatology Project (Adler et al., 2003), and to verify mean sea level pressure and geopotential height, we use ERA-Interim reanalysis, for consistency with the target data set for nudging. Note our results are insensitive to using ERA5 (Hersbach et al., 2020) instead of ERA-Interim. The NAO index is constructed using the principal component of the leading empirical orthogonal function of mean sea level pressure in the North Atlantic domain as defined in Hurrell (1995).

Link Between Tropical Rainfall and Sea Ice
We first motivate our study by exploring the link between observed tropical rainfall variability in the Pacific during early autumn and BK sea ice variability during late autumn. We compute average monthly rainfall in two boxes in the tropical West Pacific (5°S-25°N, 110°E-140°E) and tropical East Pacific (5°S-10°N, 160°E-27°0E), as defined in Scaife et al. (2017) to capture coherent regions of rainfall variability. We compute monthly BK sea ice extent, by summing the area of all grid boxes that contain sea ice concentration greater than 15%, within 65°N-85°N, 10°E-100°E. We first construct a multiple linear regression (MLR) between September-October averaged rainfall in the two boxes and use this to predict October-November averaged BK sea ice (Figure 1a). We find that the MLR can explain a significant proportion (∼25%, r = 0.5) of the variance in observed late autumn BK sea ice, suggesting an important link between the tropical Pacific and BK sea ice variability during this 37-year period in observations. This link is supported by the physical mechanisms proposed in Lee (2012) and Henderson et al. (2014) who suggest both ENSO and MJO variability can excite poleward propagating Rossby waves into the extra-tropics, which subsequently impact Arctic sea ice through poleward heat fluxes. For the ensemble percentile ranges, a member is randomly selected each year to create a comparable timeseries to observations, and an MLR is constructed. This is repeated 10,000 times, and both the 5, 10, 90, and 95th percentile of correlations is shown by the bars. This process is repeated including the year 1997 and excluding the year 1997.
If we reconstruct the MLR using the restricted period 1993-2015 inclusive, to match the model experiments here, we find the observed correlation remains strong (r = 0.43; Figure 1b). We compare this directly with the model ensemble, by randomly selecting an ensemble member for each year and recalculating the MLR with model tropical rainfall and model sea ice. We repeat this 10,000 times to produce a robust confidence interval for the model ensemble ( Figure 1b). This reveals that the observed correlation lies within the range simulated, suggesting the model may be able to capture a link between tropical rainfall and BK sea ice of comparable magnitude to that observed, but note that the observed correlation lies in the upper end of the model distribution. We also consider the forced response to tropical variability in the model, using the ensemble mean in the MLR. We compute sea ice extent in each member respectively before averaging across members. We find that the ensemble mean shows that around 20% of the modeled variance in BK sea ice is explainable by the predicted ensemble mean rainfall (r = 0.4). However, the above results depend on the inclusion of the very strong El Niño of 1997. We therefore repeat the analysis after excluding the year 1997. If 1997 is excluded from the timeseries, the observed correlation lies outside the 90% confidence interval from the model (Figure 1b), suggesting a potentially weaker link in the model than the observations. We revisit the response in 1997 later.

Reproducing the Observed Mid-High Latitude Circulation in the Model
We next further determine if our model simulations can reproduce observed interannual variability in BK sea ice through prescribing the entire tropical state (Figure 2a). Given persistence in sea ice during autumn, we compute the BK sea ice anomaly in October-November relative to September of that year and determine the correlation between these changes in model ensemble mean and observed sea ice. This provides a fair comparison as all model simulations are initialized with observed sea ice on the 1 September 2000 and observations contain considerable interannual variability in sea ice during September. We find a weak correlation (r = 0.3; p = 0.1) between the modeled and observed changes in sea ice, suggesting that the model can reproduce some of the observed interannual variability directly from the prescribed observed tropical state in our simulations. We note that our simple method of calculating anomalies relative to the September mean does not consider potentially important factors such as non-linear air-sea interactions that depend on the initial amount of sea ice, which may partly explain the weak model signal here. We discuss these results later in the context of the signal-to-noise ratio.
We next explore the model's ability to reconstruct observed interannual variability in the late autumn NAO and wider extratropical circulation in Figures 2b and 2c. This reveals that the model can skillfully reproduce a significant fraction of the observed late autumn NAO variability (r = 0.4; p < 0.05), as found in other studies such as Hansen et al. (2017). Note that the correlation between modeled and observed changes in sea ice is statistically indistinguishable from the correlation between modeled and observed NAO. We now extend this analysis to examine how well the interannual variability in the wider extratropical band is reproduced, calculating the correlation between the model ensemble-mean and observations for each grid point (Figure 2c). We find that the model can reproduce some mid-latitude interannual variability in both the Atlantic and Pacific storm track regions from tropical forcing, but it cannot reproduce interannual variability at high latitudes, or over North America and Eurasia. This limitation of tropical influence into the high latitudes was also seen in numerical weather prediction nudging experiments by Jung et al. (2014) during winter.
We explore the link between the NAO and BK sea ice through lead-lag analysis (Figure 3), given the Atlantic has been suggested as a pathway for tropical influence on the Arctic (Luo et al., 2023). First, we find that the strong correlation in the ensemble mean between sea ice and the NAO is only present in the model when the NAO precedes sea ice (r = −0.5). This is supported by studies showing that sea ice responds to atmospheric variability as discussed earlier, and that sea ice is responding to wind anomalies and heat transport mediated by the NAO (e.g., King & García-Serrano, 2016). We also find that there is large internal variability within the model ensemble; the 5-95% confidence interval of ensemble correlations extend from −0.1 to −0.7 when the October NAO leads November BK sea ice (Figure 3). Notably, the observed relationship is just outside the 5th percentile of the model distribution, suggesting that the model rarely produces the weak interannual correlation between the NAO and BK sea ice found in observations. This weak correlation found in observations, which is insensitive to detrending or accounting for persistence by computing the sea ice anomaly, suggests either the observed correlation is particularly weak during the period 1993-2015, or that the model overestimates the strength of the atmosphere to sea ice relationship, or perhaps that some feedback of the ice onto the NAO is too weak. We note that Karpechko et al. (2015) found similar results for Baltic sea-ice. The large internal variability of the link between the NAO and BK sea ice during late autumn again necessitates the use of large ensembles to detect a signal.

Tropical Influence During a Strong El Niño Autumn
We investigate the strong El Niño year of 1997 in observations and the model ensemble ( Figure 4). This event was one of the strongest on record, with SSTs over 4K above average in the East Pacific (McPhaden, 1999). In Figure 4a, we show 200 hPa eddy stream-function over the northern hemisphere to identify anomalous stationary wave patterns in late autumn 1997. We find statistically significant observed anomalies over the Atlantic and the Pacific sectors in the model, and also at higher latitudes in north-east Asia (Figure 4a). We note that the high latitude pattern found in observations is largely reproduced by the ensemble mean, suggesting that the model is capable of reproducing the same teleconnection found in observations, albeit with weaker amplitude (Figure 4d). This pattern projects onto the negative NAO and the Pacific North-American pattern in both the model ensemble mean and observations (Figures 4b and 4e). Additionally, we see significant sea-ice loss in the Barents and Kara seas in the model, as occurred in observations for autumn 1997. At least in the year of 1997, the model is able to faithfully capture the tropical-high latitude teleconnection pattern during autumn and its impact on BK sea ice. This shows that when the forcing is very strong, tropical signals can reach the high latitudes and affect sea ice and that they are discernible from internal variability in this case.

Conclusions
Autumn sea ice is commonly related to winter weather across the extra-tropics, yet remote drivers of its variability have not been extensively explored. In this study, we explore tropical drivers of interannual BK ice variability through nudging experiments. Our results show that the model can reproduce observed patterns of high-latitude circulation and sea ice in late autumn during strong tropical forcing, such as the El Niño year of 1997, and thus the tropics may be a source of predictability of late autumn sea ice. When considering the full 23-year period, we find the model weakly reproduces interannual variability of BK sea ice (r = 0.3), and the observed connections between the tropics and Arctic sea ice are in the upper end of the distribution of modeled cases. This suggests a possible underestimation of the link in models, consistent with known signal-to-noise errors in the simulated NAO in models (Scaife & Smith, 2018).
We explored components of a possible teleconnection linking the tropics to BK sea ice via the NAO to try and reconcile why the model simulations are only able to reconstruct some of the observed interannual variability in BK sea ice. First, we find that the model can skillfully reproduce some of the interannual variability in the NAO, indicating a causal link between the two. Second, we find a strong link between the NAO and BK sea ice when the NAO leads BK sea ice in the late autumn in the model. However, we note that there is considerable interannual variability in this relationship within the model, and the observed link lies on the edge of 5/95th percentile range in the ensemble. The ability of the model to reconstruct the high latitude anomalous pattern in the year 1997 and skillful reconstruction of MSLP variability in both the Atlantic and Pacific basins supports the existence of a causal teleconnection between the tropics and Arctic in autumn.
Our findings suggest two plausible explanations that may explain why the model is only able to weakly reconstruct observed sea ice variability through tropical forcing alone. Our first hypothesis is that a low signal-to-noise ratio in the extra-tropics may be hampering any detectable impact on Arctic sea ice outside of years of strong tropical forcing, such as the year 1997. This may in part be due to our experiment design, as initializing with the background state of autumn 2000 may dampen planetary wave activity through destructive interference. The low signal-to-noise ratio in the extratropics is well known ( (Figure 2b) and in the link between the NAO and BK sea ice in the model (Figure 3). To overcome this, large ensembles are required , and thus our limited sample of 20 members may be insufficient for detecting teleconnections to the extra-tropics. An alternative hypothesis that may explain the model results here is that the observed correlation between tropical rainfall and BK sea ice may be spuriously strong within this time period. Prior work showed that during winter, both internal and tropically forced variability could explain the link between autumn Arctic sea ice and the winter NAO (Warner et al., 2020). In this work, we have shown that autumn NAO variability can be directly attributed to tropically forced variability but only if the tropical forcing is strong, and Figure 1 indicates that although it is just possible for model ensemble members to simulate such a strong link, this is unlikely (observations lie at the 95th percentile of the ensemble). The low signal-to-noise ratio found within the model NAO and BK sea ice within the model ensemble suggests that larger ensembles may be required to isolate interannual variability during years where tropical forcing is smaller. We recommend that large ensembles are used to infer tropical impacts on the high latitudes in light of this signal-to-noise issue.