Contributions from intrinsic low-frequency climate variability to the accelerated decline in Arctic sea ice in recent decades

Abstract. In recent decades, the Arctic sea ice has been declining at a rapid pace as the Arctic is warmed at a rate of twice the global average. The underlying physical mechanisms for the Arctic warming and accelerated sea ice retreat are not fully understood. In this study, we apply a relatively novel statistical method called Self-Organizing Maps (SOM) to examine the trend and variability of autumn Arctic sea ice in the past four decades and their relationships to large-scale atmospheric circulation changes. Our results show a large portion of the autumn Arctic sea ice decline between 1979 and 2016 may be associated with anomalous autumn Arctic intrinsic atmospheric modes. The Arctic atmospheric circulation anomalies associated with anomalous sea surface temperature patterns over the North Pacific and North Atlantic influence Arctic sea ice primarily through anomalous temperature and water vapor advection and associated radiative feedback.

network-based method, SOM uses unsupervised learning to determine generalized patterns in complex data. The technique can reduce multidimensional data into two-dimensional array consisting of a matrix of nodes. Each node in the array has a reference vector that displays a spatial pattern of the input data. All 70 patterns in the two-dimensional array represent the full continuum of states in the input data. The SOM algorithm also is a clustering technique, but unlike other clustering techniques, it does not need a priori decisions on data distribution. Unlike the EOF analysis, the SOM technique does not require the orthogonality of two spatial patterns. A detailed description of the SOM algorithm is given in Kohonen (2001). 75 In this study, the SOM technique was used to categorize anomalous seasonal sea ice concentration patterns north of 50°N for autumn (September-October-November). The autumn seasonal anomalies are calculated by subtracting the climatology, which is the overall mean for all 38 autumns in the study period 1979-2016, from the autumn mean for each year. The anomalous sea ice pattern for each autumn is assigned to the best-matching SOM pattern on the basis of minimum Euclidean distance. Pattern 80 correlations between anomalous sea ice field for each autumn and its corresponding best-matching SOM pattern are used to determine the number of SOM nodes or grids (Lee and Feldstein, 2013). We calculate spatial correlation coefficients for different number of SOM grids ranging from 2×2 to 4×5 (Table 1).
There is a large increase in correlation from 2×4 to 3×3 and thus the 3×3 SOM grid is chosen for the analysis. Smaller grids may not adequately capture the variability of autumn Arctic sea ice, whereas 85 larger grids, although providing more details, do not alter the results and conclusions. The contribution of each SOM pattern to trends in Arctic sea ice concentration is calculated by the product of each SOM pattern and its frequency trend (Johnson, 2013), where frequency is calculated by the number of the occurrences of each SOM pattern divided by the total number of autumns over the study period (38) and the trend of the frequency time series for each SOM pattern is determined through linear regression. The 90 sum of the contributions from all SOM patterns represents the trends in Arctic sea ice explained by the SOM patterns, which in this case indicates trends resulting from low-frequency variability. The significance of the trends in the time series for each SOM pattern is tested using the Student's t-test.

SOM results 95
Sea ice concentration anomalies occur mainly in the marginal seas from the Barents Sea to the Beaufort Sea, with maximum anomalies in the Barents Sea and the Kara Sea, and over the Beaufort Sea and the We examine trends in the frequency of occurrence for each SOM pattern and their contribution to trends in the Arctic autumn sea ice concentration. Figure 2 shows the occurrence time series for each SOM 110 pattern. The nodes with spatially uniform changes appear to be separated into two clusters with those showing all positive anomalies (Nodes 5, 6, 8 and 9) appearing in the 80s and 90s and those having all negative anomalies (Nodes 1 and 4) appearing after 2000. The transition from all positive anomalies in the earlier part, to all negative anomalies in the later part, of the time series is consistent with the trends in the observed Arctic sea ice concentration during the same time period. Not surprisingly, the transition 115 appears to be dominated by the two strongest and most frequent patterns denoted by Node 9 (all positive and occurring from the 1980s through mid 1990s) and Node 1 (all negative and occurring exclusively after 2005). Only these two nodes have linear trends that are statistically significant at above 95% confidence level. The slopes of the trend lines for these two nodes are opposite but the values are similar (0.027 yr -1 for Node 1 and -0.021 yr -1 for Node 9). The other nodes have statistically insignificant trends 120 with magnitudes less than 0.01 yr -1 . in the autumn sea ice loss (Fig. 4b). Among them Node 1 explains the largest portion (33%) of the total trend, followed by Node 9 (21%), with the other 7 nodes together accounting for only 6% of the loss. 125

Potential Mechanisms
To explain the spatial patterns depicted by the two dominant nodes, we made composite maps over the years of occurrences for Nodes 1 and 9, respectively, and Figs. 5-7 show composite patterns of the anomalous SST and anomalous atmospheric circulations represented by the 500-hPa geopotential height, 850-hPa wind, surface to 750-hPa specific humidity, surface downward longwave radiation, and surface 130 air temperature. For Node 1, the SST composite pattern resembles negative phase of the PDO in the North Pacific and positive phase of the AMO in the North Atlantic (Fig. 5). The positive SST anomalies over the mid-latitude North Pacific produce local negative upper-tropospheric vorticity anomalies (Hoskins and Karoly, 1981) which excite a wave train with zonal wave number two that propagates eastward to North America, North Atlantic, and Eurasian Continent (Fig. 5). Over the Arctic, the pattern 135 resembles negative phase AD, with a center of negative 500-hPa height anomalies over Greenland and the Baffin Bay, and a center of positive anomalies over the Kara Sea. The zonal pressure gradient between the two centers induces anomalous low-troposphere southwesterly and southerly winds over the North Atlantic Ocean (Fig. 6) that transport warm and moist air from North Atlantic into the Arctic Ocean north of Eurasia, thus increasing surface air temperature and humidity and reducing sea ice 140 concentration in the Arctic (Fig. 7). The higher moisture content in the Arctic surface air also facilitates the occurrence of water vapour and cloud radiative feedback process (Sedlar et al., 2011) during which increased downward longwave radiation (Fig. 7)  the North Atlantic Ocean induced by anomalous Icelandic low are unfavourable for warm air intrusion into the Arctic Ocean (Fig. 6), a result also indicated in a previous study (Kim et al., 2017). Despite the occurrences of anomalous southwesterly winds over the Barents Sea and southeasterly winds over the 155 Sea of Okhotsk, the cold advection from anomalous cold sea water prevents the Arctic sea ice from melting. Meanwhile the cold advection also reduces water vapor content in the lower troposphere and the resulting smaller downward longwave radiation facilitates the occurrence of negative surface air temperature and positive Arctic sea ice anomalies north of Eurasia (Fig. 7). The northerly winds over the Beaufort Sea, northeastern Canada and Greenland Sea also contribute to the decrease in surface air 160 temperature and increase in the sea ice concentration.
The opposite patterns in Nodes 1 and 9 may be explained by the differences in the water vapor-radiation feedback process resulting from anomalous temperature and especially water vapour transport by anomalous atmospheric circulations associated with different patterns of SST anomalies over the North Pacific and Atlantic. The patterns of SST anomalies are nearly symmetric for the two nodes over the 165 North Atlantic, but they are somewhat asymmetric over the North Pacific. For Node 1, there is one center of high SST over the central North Pacific, whereas Node 9 is associated with two centers of low SST: one over the Coast of Japan and another in central North Pacific. These differences in the SST anomaly patterns lead to different wave trains and high-latitude atmospheric circulations.  1980-1982, 1986-1989, 1992, and 1996. With original sea ice time series, higher and lower sea ice periods are usually considered as before and after the late 1990s. The SOM-based composites allow for better depiction of atmospheric circulation patterns that have significant impact on sea ice trends. 175

Conclusions
We investigate the potential mechanisms for the autumn Arctic sea ice decline for the period 1979-2016 using the SOM method. Our results show that a large portion of the autumn Arctic sea ice loss may be associated with the changes in the temperature and water vapour transport and the associated water 180 vapour radiation feedback resulting from anomalous atmospheric circulations linked to SST anomalies Atmos. Chem. Phys. Discuss., https://doi.org/10.5194/acp-2018-127 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 4 April 2018 c Author(s) 2018. CC BY 4.0 License. over the North Pacific and North Atlantic. This result provides further evidence that the mid-latitude SST anomalies play a vital role in the accelerated Arctic sea ice decline in recent decades (Ding et al., 2017;Yu et al., 2017). An important finding is that the opposite pattern of the Arctic sea ice anomalies during the early (positive) and later (negative) parts of the 1979-2016 period is not associated with opposite 185 phase of an atmospheric circulation mode; but instead the change may be explained by two different atmospheric circulation patterns (AO and AD) associated with an asymmetry in the anomalous SST distributions over the North Pacific. The teleconnections between the Arctic sea ice variability and mid-latitude SST anomalies suggest that on a decadal or longer time scale it may be necessary to include the Arctic sea ice and mid-latitude SST interactions or feedbacks in any investigations of Arctic warming 190 and sea ice decline and their potential influence on mid-latitude weather and climate, an area of active research in recent years (Barnes and Screen, 2015;Overland and Wang, 2015;Francis and Skific, 2015).
Finally, the results here help highlight the large contributions from the decadal-scale natural climate variability to Arctic climate change, though further studies using coupled global atmosphere-ocean-sea ice models are necessary to fully understand the physical mechanisms. 195

Data Availability
All data used in the current study are publicly available. The monthly sea ice concentration data are available from the National Snow and Ice Data Center (NSIDC) (http://nsidc.org/data/NSIDC-0051), the ERA-Interim reanalysis data are available from the European Center for Mid-Range Weather Forecasting (https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era-interim) and the SST 200 data are available from the Hadley Centre for Climate Prediction and Research (http://www.metoffice.gov.uk/hadobs/hadisst/).

Competing interests
The authors declare that they have no conflict of interest.

Author Contributions 205
L. Yu designed the study, with input from S. Zhong, and carried out the analyses. L. Yu and S. Zhong prepared the manuscript.