Introduction

The Arctic has undergone dramatic changes in temperature and in sea ice extent since routine measurements have begun, with temperatures warming at more than double the rate of the global average1,2,3,4 and September sea ice extent (SIE) declining by over 40%5. Greater warming in the Arctic than surrounding lower latitudes, referred to as Arctic amplification, can largely be explained by the influence of declining sea ice on atmospheric radiative feedbacks. The presence of sea ice elevates atmospheric stability and inhibits the increase in longwave energy radiated back to space in response to surface warming compared to ice-free regions6. In addition, absorption of solar radiation at the surface increases when snow and ice retreat7,8,9,10,11,12. From October - January, when incoming solar radiation is largely absent, the increasing surface air temperature trends are largely associated with the increase in the downward longwave radiation13. However, even though global climate models incorporate such feedback processes, ensemble projections of sea ice decline tend to be quite conservative with respect to the observed trends14,15,16,17,18. Recently, evidence has been mounting on the impact that Arctic cyclones can have on sea ice at very short time scales19,20,21,22,23,24,25,26,27. The present study will establish that Arctic cyclones associated with sea ice loss at short time scales occur in areas of relatively thin sea ice where there are enhanced pressure gradients. Though the exact mechanisms of how an Arctic cyclone can rapidly change sea ice are incompletely understood, some studies suggest that ocean waves induced by the wind of an Arctic cyclone on the surface can break up sea ice under certain conditions18,28,29,30,31,32,33,34. Even when Arctic cyclones are well-represented in numerical models35,36,37, the sea ice dynamics surrounding upper-ocean mixing and the breakup due to waves is a missing process in most models, resulting in little skill in sea ice reductions due to Arctic cyclones18. Notwithstanding sea ice reductions alone, there is growing evidence that a warming Arctic can impact large-scale atmospheric circulation patterns throughout the Northern Hemisphere38,39,40,41,42,43,44,45,46,47,48, thus gaining knowledge of the variability in sea ice due to short time-scale features such as Arctic cyclones is important in order to assess the global risks of climate change. This study investigates the relatively short, synoptic-timescale reductions in sea ice, which we call very rapid sea ice loss events (VRILEs), and whether there are common atmospheric features associated with VRILEs such as Arctic cyclones.

The term RILE49, for rapid ice loss event, describes the very large, intermittent trends of about 5-yr duration in September SIE, first discovered in global climate models50 and seen in the Arctic over 2007–2012. Back-to-back years with an enhanced frequency of VRILEs could lead to a RILE. A fundamental understanding of how Arctic cyclones could cause VRILEs is lacking and VRILEs have not been predicted more than a week or two in advance51,52. The timing of Arctic cyclones and the state of the underlying sea ice are important factors that influence the sea ice response27,53,54. When a surface cyclone develops early in summer, it may act to preserve sea ice by increasing cloud cover and decreasing the surface absorbed solar radiation25,55. In contrast, while most surface cyclones do not lead to ice loss25,26,56, it may be possible for a surface cyclone in late summer to enhance melt or inhibit growth by mixing up heat from the upper ocean’s near-surface temperature maximum that may otherwise be sequestered under a freshwater capping layer21,27. Cyclone-driven ocean surface waves can propagate further into thinner sea ice, accelerating sea ice floe breakup29,30,34,57,58. The region subject to wave propagation, known as the Marginal Ice Zone (MIZ), with lower sea ice concentration rimming the more concentrated pack ice, has widened in summer in recent decades59.

Results and discussion

Arctic cyclones

One source of short timescale variability in SIE are atmospheric forcings associated with synoptic-scale Arctic cyclones60. Arctic cyclones can have radii of up to  ~2000 km61, lifetimes as long as several weeks23, are frequently located in the central Arctic Ocean during the summer62, and often form in association with tropopause polar vortices (TPVs)19,23,63,64,65,66,67,68,69, when an upper-level potential vorticity anomaly moves over a lower-level temperature gradient63,70. TPVs are generally sub-synoptic scale features on the tropopause that can be present days to months before an Arctic cyclone forms71 and thus could offer insight into the physical processes that are important for VRILEs. An example of an Arctic cyclone and TPV (8- and 85-day lifetimes, respectively) in August 2006 is shown in Fig. 1a, colocated in a region of reduced sea ice near the MIZ of the North Atlantic (Fig. 1b). This area of reduced sea ice was not present before the cyclone, suggesting the possibility that ocean waves and/or atmospheric wind from the cyclone may have played an important role in the rapid loss of sea ice. While this is just one example of an Arctic cyclone and TPV associated with an apparent VRILE, we will evaluate the significance, frequency, and characteristics of short timescale variability in Arctic SIE under the hypothesis that Arctic cyclones and TPVs are a common association to VRILEs.

Fig. 1: Observational-based analyses of reduced Arctic sea ice underneath an Arctic cyclone and tropopause polar vortex.
figure 1

a Potential temperature on the dynamic tropopause (colors; color interval 5 K) and sea level pressure (black contours; contour interval 4 hPa; only values at or below 1004 hPa are shown) at 00 UTC 21 August 2006 and b sea ice concentrations (colors) from passive microwave satellite radiometry composited over 21 August 2006. The red rectangle highlights a region of locally low sea ice concentration (b) under an observationally-based reanalysis of a TPV and Arctic cyclone (a). The dynamic tropopause is the 2 PVU surface, where 1 PVU = 106 K m2 kg−1 s−1. The sea ice concentration points near the pole (latitudes ≥ 87.5°N) are masked out.

Full size image

Short timescale variability in Arctic sea ice

Observations indicate that variability in SIE (Fig. 2a) and in moving 3-day sums in daily changes in Arctic sea ice extent (ΔSIE) (Fig. 2b) is substantial not only for intraseasonal and seasonal timescales, but also for shorter timescales of about 18 days or less. Since changes in SIE can, to some extent, be persistent, we expect that ΔSIE is less susceptible to persistence and more to atmospheric forcing; hence, we subsequently choose to focus on ΔSIE for the remainder of this analysis. The observation shown here that ΔSIE variability is significant at timescales of 18 days or less motivates an analysis of ΔSIE that further isolates these shorter timescales as described below. While global climate model projections from the Community Earth System Model-Large Ensemble (CESM-LE)72 reproduce variability in ΔSIE quite well for periods of over 40 days, variability in ΔSIE is substantially lower than observations for shorter timescales18. This could mean VRILEs may not be a significant contribution to ΔSIE in CESM-LE and that the CESM-LE, and potentially other climate models, are not representing the physical processes that produce VRILEs. Furthermore, observed atmospheric variability from intraseasonal variations such as the Arctic Oscillation (AO) and North Atlantic Oscillation are significant at timescales of as short as 1 week (not shown), and since ΔSIE is in part associated with these patterns of atmospheric variability40,73,74,75, CESM-LE may also not be representing the full range of impacts from them.

Fig. 2: There is significant sea ice variability at high frequencies (weather timescales) that is not represented in climate models.
figure 2

a SIE and b ΔSIE power spectral density (PSD) for a random CESM-LE ensemble member projection of the near present day equivalent period (green) compared to observations from the NSIDC sea ice index (1989–2023; blue). The 95% confidence interval from the theoretical red noise spectrum (dashed red) is also shown, and thus PSD values above this dashed red line indicate statistically significant a SIE or b ΔSIE variability. The vertical dashed black line denotes the 18-day cutoff period used in the high bandpass filter to retain the weather timescales. Units of PSD here are (variance Hz−1).

Full size image

VRILEs occur throughout the year, but have the largest sea ice reduction during the warm season of June, July, and August (JJA), when they also exhibit an increasing frequency in the most recent decade, and partially contribute to the increasing variability in September SIE (Fig. 3). The largest sea ice reductions occur in June and July, when each VRILE consists of around 0.3–0.7 × 106 km2 of SIE reduction (Fig. 3a). From September to February, when there is typically expansion in SIE due to freezing, VRILEs occur when there is simply less expansion in SIE rather than a reduction in SIE. In contrast, warm-season VRILEs can have a long-lasting impact on SIE since VRILEs are typically associated with 0.2–0.3 × 106 km2 more ice loss than the typical background SIE reduction. VRILE frequencies usually consist of a cold-season peak between October and February and a warm-season peak in July or August (Fig. 3b). As will be described in further detail in “Methods”, we separate VRILEs into two types based on the method used to calculate them. The first method (ΔSIEmean removed) are extreme ΔSIE events based on anomalies with respect to the climatological mean, while the second method (ΔSIEbwfilt) isolates the events associated with only the short timescales that have significant variability in observations. ΔSIEmean removed (containing daily values) exhibits a statistically significant positive trend over time during the summer (Table 1), with the largest increases in frequency during the summer of the most recent decade (Fig. 3c and Table 2). This trend in ΔSIEmean removed is not surprising, given the mean SIE of the most recent decade is much lower than previous decades76. Since ΔSIEbwfilt can be attributed to high-frequency variability, there are considerably fewer events than ΔSIEmean removed with no long-term trends since these longer-term variations have been removed by the Butterworth Filter. In 2012, six of the twelve summer (June-August) VRILEs occurred at the same time as the Great Arctic cyclone19, while there are six ΔSIEmean removed events in the “preconditioning” period in which a persistent anticyclone promoted warm temperatures and hence a prolonged period favoring ice melt in June53.

Fig. 3: Very rapid sea ice loss events contribute to an acceleration of Arctic sea ice reduction.
figure 3

a The 30-yr climatological range of ΔSIE (gray) compared with the bottom 5th percentile mean-removed VRILEs (ΔSIEmean removed; red) and range of filtered VRILEs (ΔSIEbwfilt; blue). Distribution of the (b) number of VRILEs for decadal periods with 1991 through 2001 (black), 2002 through 2012 (blue), and 2013 through 2023 (red) by month and (c) total number of filtered VRILEs (blue) plus mean-removed VRILEs (red) in June-August (JJA) by year. d 1 September SIE (black) and 1 September SIE minus the accumulated sea ice reductions from ΔSIEtotal (gray). The fractional reduction in SIE from ΔSIEtotal (red) is shown for comparison using Equation (1) and expressed as a percentage. The number of VRILEs for decadal periods 1991 through 2001 (black), 2002 through 2012 (blue), and 2013 through 2023 (red) by ΔSIE for (e) JJA and (f) all months except JJA.

Full size image
Table 1 There is a significant trend over time in summer extreme changes in sea ice extent events based on anomalies with respect to the climatological mean
Full size table
Table 2 Trends over time in summer extreme changes in sea ice extent events based on anomalies with respect to the climatological mean are largest in the most recent decade
Full size table

The accumulated reduction in warm season SIE from VRILEs contributes to as much as  ~1 × 106 km2 of the seasonal loss and is significantly increasing in time at a rate of 3.1% per decade (p-value of 1.78 × 10−3), with over 20% of the total warm season reduction in SIE in 2012, 2018, and 2019 occurring from all ΔSIE events (Fig. 3d). The pattern of interannual variability in SIE does not change much in consideration of ΔSIE, implying that large year-to-year differences in SIE can not be attributed to VRILEs alone. The ΔSIE per VRILE is also increasing during the warm season, especially in the extreme loss categories of 0.5–0.6 × 106 km2 of sea ice loss per event (Fig. 3e), and is similar to those outside of the summer season (Fig. 3f).

Spatial characteristics of very rapid sea ice loss events

We next show there are common atmospheric patterns associated with VRILEs, hypothesizing that sufficiently thin sea ice in summer is systematically associated with strong atmospheric winds since winds can potentially promote accelerations in ice loss through processes such as breakup of ice from ocean waves or ocean upwelling. The 1st-year ice age extent contour is used as an approximate measure of the relatively thin vs. thick sea ice since sea ice that is less than 1 year old is at least thin enough to succumb to seasonal forcing.

On average, VRILEs occur in regions of relatively strong low-level atmospheric pressure gradients in combination with thinner, younger sea ice (Fig. 4). During the warm season, the combined low-level pressure gradient and thin sea ice occurs over a broad region of the Arctic Ocean centered over the northern Beaufort Sea between the Beaufort High Pressure and Arctic cyclones (Fig. 4a, c), while the rest of the year they are mostly in the 1st-year ice of the MIZs of the North Atlantic and North Pacific between the Beaufort High Pressure and cyclones in the midlatitude storm track (Fig. 4b, d). Note that Hudson Bay has been becoming increasingly more ice free during the warm season, leading to fewer VRILEs over time (compare Figs. 4a, c). Outside of the warm season, there are a high number of VRILEs in the Sea of Okhotsk and the Bering Sea, also in locations where sea-level pressure gradients are enhanced between the synoptic-scale cyclones in the North Pacific storm track and the Beaufort High along the MIZ. Indeed, when averaged relative to the location of the VRILE, it is clearly evident that the VRILE location is centered in a region with an enhanced pressure gradient between a surface cyclone and anticyclone (Fig. 4e). On average, VRILEs are associated with a nearby surface cyclone and tropopause polar vortex (TPV) (Fig. 4f), and all surface cyclone cases are associated with an antecedent TPV (not shown). These results are consistent with a previous study, where it was found that nearly all very-intense cyclones were related to a TPV64. On average, the surface cyclone and TPV exhibit a tilt with height at the time of the VRILEs, which indicates the surface cyclones are generally strengthening at the time of the VRILE.

Fig. 4: Very rapid sea ice loss events occur in regions of strong pressure differences between cyclones and the Beaufort High.
figure 4

a–d Locations of filtered VRILEs, computed with (ΔSIEbwfilt), and e, f June-August 1989–2023 composite tropopause potential temperature (colors) and sea level pressure (contours) standardized anomalies. Filtered VRILE locations in a, c are for June–August while b, d are for non June–August months for the years a, b 1989–1998 and c, d 2014–2023. VRILE locations are determined by the number of sea ice concentration loss objects at any 0.5° grid point. In (a–d), mean sea level pressure is shown by gray contours with a 1 hPa interval and thick (thin) cyan contours are the monthly median (maximum) of a, c June–August and b, d non June–August 1-year ice age extent from the a, b 1989–1998 median (maximum) and c, d 2014–2023 median (maximum) ice age extent. Composites in e, f are averaged relative to the e VRILE location and f closest Arctic cyclone with dashed (solid) contours indicating negative (positive) anomalies.

Full size image

To conclude, SIE has been objectively identified to exhibit significant, short, “weather”, timescale variability at periods of 18 days or less in the Arctic. The times of the extreme reductions in SIE corresponding to these short timescale variations, referred to as VRILEs, or VRILEs, are found to occur throughout the year, with a significant increase in the frequency of VRILEs since 1979 during the warm season from June-August. The locations of VRILEs are located over an extensive region of 1st-year Arctic sea ice during the warm season but is confined to the MIZs of the North Atlantic, Bering Sea, and Sea of Okhotsk outside of the warm season. Furthermore, the locations of the VRILEs are in regions of strong low-level atmospheric horizontal pressure gradients that are enhanced between Arctic cyclones and the Beaufort High. The Arctic cyclones are accompanied by an antecedent TPV. TPVs can be present days to weeks before the formation of a surface cyclone63,70,71, and yet are located in atmospheric layers that are exceptionally devoid of reliable observations, particularly moisture, which has been shown to be important for TPV maintenance and intensification77. This, combined with the relatively low forecast skill of some Arctic cyclones36, suggests that TPVs may be an important feature of interest in improving the predictability of sea ice loss at up to seasonal timescales. Note that while a number of studies have established a link between Arctic cyclones and sea ice24,25,26,27,34,78, this is the first study to establish the association to enhanced pressure gradients over thin sea ice. Thus, this study has proposed a distinction for Arctic cyclones that lead to sea ice loss versus those that do not. Furthermore, this study is the first to establish a connection to TPVs. Given these new insights, future studies are needed to address the underlying physical mechanisms causing VRILEs.

Methods

We use daily SIE from the National Snow and Ice Data Center (NSIDC) Sea Ice Index79. Before 1989, SIE is only available every second day. Given that this study focuses on day-to-day variations in SIE, we restrict the analysis to the years 1989–2023 when SIE is available daily. ΔSIE is the 3-day change in SIE, i.e, SIE(n + 1) − SIE(n − 2), where n = day. Due to sudden, erroneous jumps in SIE that occur on the first day of every month80, we omit any analysis of ΔSIE on the last 2 days and first day of each month. These sudden jumps were found to occur due to the difference in ocean masks used in the Arctic to filter residual weather effects far from the ice edge where sea ice is not possible, and these ocean masks are only updated once a month on the first day of each month.

Two different methods of quantifying ΔSIE for short-time scale variability are considered here: Extremes of all (1) ΔSIE from daily SIE anomalies with respect to a daily SIE climatological mean annual cycle for 1990–2018 (ΔSIEmean removed), and (2) spectrally filtered ΔSIE using a high-pass Butterworth Filter using a cutoff period of 18 days to isolate the short timescale variability that can generally be interpreted as “weather” (ΔSIEbwfilt). Extremes are treated as the tail of the distribution of ΔSIE loss, here the lowest 5th-percentile in ΔSIE. Method (1) can be interpreted as the extreme ΔSIE events based on the anomalies in ΔSIE with respect to the long-term climatological record. While neither method above is expected to be a perfect measure of ΔSIE, both methods have merits and shortcomings. Method (1) identifies all extreme instances of SIE loss; however, variability from processes at all timescales is implicitly included. For example, sea ice melt may be enhanced during the negative phase of the AO, when near-surface temperatures are anomalously warm and cloud cover is less likely. Even though SIE could decrease substantially over a short duration of time (days), the frequency of such “events” is relatively low (months), and method (1) can not distinguish the two scales apart. In contrast, method (2) explicitly removes variability associated with long-term climate trends, seasonal oscillations, as well as intraseasonal variations. Method (2) emphasizes the relatively short timescale variability in ΔSIE that could occur on synoptic (“weather”) timescales from atmospheric processes, including Arctic cyclones and their associated fronts. By definition, method (2) isolates the events associated with the short timescales that have significant variability in observations, and that may be missing from climate models. The union of events from both methods (1) and (2) using ΔSIEtotal are the total VRILEs. ΔSIEtotal is the sum of ΔSIEmean removed and ΔSIEbwfilt. When both methods capture the same event, the event is added individually to both ΔSIEmean removed and ΔSIEbwfilt, but only one event is added to ΔSIEtotal. The fractional SIE reduction per warm season from ΔSIEtotal is computed by comparing the daily ΔSIEtotal accumulated on all respective dates from 1 June - 31 August, and dividing by the daily ΔSIE accumulated each day from 1 June to 31 August:

$$f=\frac{\mathop{\sum }_{{{\rm{1June}}}}^{{{\rm{31August}}}}\Delta {{{\rm{SIE}}}}_{{{\rm{total}}}}}{\mathop{\sum }_{{{\rm{1June}}}}^{{{\rm{31August}}}}\Delta {{\rm{SIE}}}}.$$
(1)

Occurrences of VRILEs on consecutive days are not eliminated but are recognized as possibly being related to the same physical feature, such as a long-lived surface cyclone. To account for these occurrences, we reduce the VRILEs into a subset referred to as “Unique VRILEs”, where any series of VRILEs identified on adjacent days are counted as a single event, with the date corresponding to the last day in the series. VRILEs are increasing in frequency during the warm season, regardless of whether daily back-to-back VRILEs are counted (“Unique cases”) or if the tail of the extreme distributions is relaxed to the 10th-percentile (not shown). Trends of unique VRILES are weaker while remaining significant. Since atmospheric patterns between consecutive days are generally correlated, we include only unique VRILEs in our analysis to reduce the possibility of biasing the spatial statistics to less common, long-lived atmospheric features.

Power spectral density (PSD) is computed from the daily ΔSIE timeseries by subtracting the mean and computing the absolute value of the square of the Fast Fourier Transform coefficient. Statistical significance is established at the 95% confidence level by comparing the computed power spectra against the theoretical red noise spectrum using the lag correlation function for a first-order linear Markov process with a one-lag autocorrelation81.

ΔSIE does not contain information about the specific geographic location of a VRILE. In order to map the VRILE to a location to analyze composite patterns, we compute the change in daily NSIDC sea ice concentration (SIC) over a 5-day period ending on the day of the VRILE. The resulting changes in SIC are then separated into objects by comparing grid point changes with each of the neighboring grid points. The largest connected region of loss is then identified as the VRILE object. The single point location of the VRILE is determined by computing the center of mass where all points within the ice loss object are weighted by the absolute value of the 5-day SIC change. Atmospheric fields from the day of the VRILE are then averaged relative to the location of the VRILE using European Center for Medium Range Forecasting Reanalysis (ERA5)82. The nearest surface cyclone is found as long as it is within 1500 km of the VRILE location up to 5 days before the VRILE. The nearest TPV is used as the reference point for the TPV-relative composites. Surface cyclones and TPVs are identified by local minima in their respective fields. Composite anomalies of sea level pressure and tropopause potential temperature are computed from ERA5 with respect to the 30-year monthly mean from 1981–2010. Composite plots are oriented such that the lowest value of tropopause potential temperature in the TPV lies at the center point.