Winter storms accelerate the demise of sea ice in the Atlantic sector of the Arctic Ocean

A large retreat of sea-ice in the ‘stormy’ Atlantic Sector of the Arctic Ocean has become evident through a series of record minima for the winter maximum sea-ice extent since 2015. Results from the Norwegian young sea ICE (N-ICE2015) expedition, a five-month-long (Jan-Jun) drifting ice station in first and second year pack-ice north of Svalbard, showcase how sea-ice in this region is frequently affected by passing winter storms. Here we synthesise the interdisciplinary N-ICE2015 dataset, including independent observations of the atmosphere, snow, sea-ice, ocean, and ecosystem. We build upon recent results and illustrate the different mechanisms through which winter storms impact the coupled Arctic sea-ice system. These short-lived and episodic synoptic-scale events transport pulses of heat and moisture into the Arctic, which temporarily reduce radiative cooling and henceforth ice growth. Cumulative snowfall from each sequential storm deepens the snow pack and insulates the sea-ice, further inhibiting ice growth throughout the remaining winter season. Strong winds fracture the ice cover, enhance ocean-ice-atmosphere heat fluxes, and make the ice more susceptible to lateral melt. In conclusion, the legacy of Arctic winter storms for sea-ice and the ice-associated ecosystem in the Atlantic Sector lasts far beyond their short lifespan.


Results
The six N-ICE2015 winter storms. Cohen et al. 34 documented how the N-ICE2015 camp observed six major winter storms (M1-M6), from 20 January and 21 March 2015 (Fig. 1). Here we analyse the tracks of these storms using the University of Melbourne's cyclone identification scheme 35 (Fig. S1). The six storms entered the Arctic Ocean from the North Atlantic through the Fram Strait, between Greenland and Svalbard 34 (Fig. S1). Storms M1, M2, M4 and M6 continued northwards after passing through the Fram Strait and tracked northwest of the N-ICE2015 camp, but Storms M4 and M6 dissipated shortly after (Fig. S1). In contrast, Storms M3 and M5 followed a more eastward trajectory from Svalbard towards Franz Josef Land and the Barents and Kara Seas 34 . As a result, these storms initially passed south of the N-ICE2015 camp (Fig. S1). Storm M5 followed the most southerly trajectory, reaching a northernmost latitude of 78.3°N. All other storms tracked north of 84°N. Over the following sections, we summarise the common patterns and characteristics of these six Arctic winter storms. However, we note that the overall impact of an individual storm is unique and will be dependent on the storm's track, strength and lifespan.   34 , Kayser et al. 36 , and Walden et al. 37 . snow from autumn and winter storms limits ice growth. The thermal conductivity of snow is an order of magnitude lower than that of sea ice 41 , and therefore snow effectively insulates the ocean from the cold atmosphere and reduces sea ice growth 29,41 . At the start of the N-ICE2015 field campaign, a 0.4-0.6 m deep snowpack was observed on first and second year ice floes 38,42,43 (Fig. 3b,c). This is double the widely used snow-on-sea-ice climatology for January in this region 44 . The disparity in snow depth from the climatology likely reflects the lack of representative early winter observations from this region 28,[42][43][44] . However, it is also possible that the unexpectedly deep snow pack is the result of more frequent and more intense winter storms bringing more precipitation to this region in recent years 3,45 . All snowfall greater than trace amounts during the N-ICE2015 www.nature.com/scientificreports www.nature.com/scientificreports/ winter was associated with a storm 34 (Fig. 2f). Studies analysing the structure of the N-ICE2015 snowpack and atmospheric reanalyses, concluded that the snowpack deepened during a series of storms in late autumn and early winter 32,46 .
The extensive N-ICE2015 observations revealed minimal (<0.1 m) thermodynamic growth on the underside of snow-covered first and second year sea ice between January and March 2015, despite prolonged periods with air temperatures below −30 °C 38,42 (Fig. 3b,c). In contrast, comparably low air temperatures during SHEBA contributed to substantial (0.2 m) winter sea ice growth when the snow pack was only 0.25 m deep, even though the ice was more than 0.5 m thicker compared to N-ICE2015 31,39 . However, ice growth rates are also controlled by many factors other than snow depth 12 .
Here we isolate the impact of snow insulation on ice growth rates during the N-ICE2015 winter using a 1-D sea ice model (Methods). Specifically, we perform two simulations. In the first simulation, we force the model with the total precipitation from an atmospheric reanalysis. This produces a 0.4 m snowpack in January, which is consistent with N-ICE2015 observations 43,46 (Fig. 4). The total thermodynamic ice growth for this multi-year ice floe is approximately 0.10 m over the entire winter season (Fig. 4). For the second simulation, we reduce the total precipitation by a factor of two, but otherwise keep the model setup identical. The lower precipitation produces a 0.25 m snowpack in January that is more consistent with the Warren et al. 44 climatology for the region. With this thinner snowpack, thermodynamic winter ice growth increases to approximately 0.35 m (Fig. 4). Hence, these simulations indicate that the insulation provided by the deep snowpack during N-ICE2015 likely reduced thermodynamic winter ice growth by 0.25 m, relative to a climatological snow depth 44 .
Several recent studies have highlighted the role of increased atmospheric heat and moisture transport, by more frequent storms, for reducing winter sea ice growth in the Atlantic Sector of the Arctic Ocean 4,10,11,13,39 . However, the extreme transient warming events during N-ICE2015 winter storms had limited impact on growth rates of snow covered sea ice (Fig. 3). Instead, we emphasise that more frequent winter storms likely contribute to more precipitation and a deeper snowpack. Our model simulations and the in-situ winter observations from N-ICE2015 demonstrate that the additional insulation provided by a deep snowpack can have a larger impact on winter ice growth than any transient atmospheric warming during storms. This is because a deep snowpack effectively decouples the ice from the atmosphere. Downward directed conductive heat fluxes propagated by warm air during storms dissipate before they reach the ice-water interface and likewise, snow insulation prevents effective upward directed fluxes to cool the ocean surface during cold quiescent period between the storms, resulting in minimal ice growth rates (Fig. 3d). While the crucial role of snow insulation on winter ice growth is well known 31,41,47 , the consequences of snow deposited on sea ice during winter storms for Arctic sea ice have largely been overlooked. This is due to a lack of in-situ observations and large uncertainties in precipitation and snow depths from models, reanalyses, and remote sensing products 29 . Interestingly, we note here that the impact of snow insulation on ice growth is known to be largest for snow falling on thin sea ice early in the winter season 46 , and the main increase in winter storm activity from 1979-2016 is reported during the early winter 3 .
Winter storms promote snow-ice formation. While thermodynamic ice growth was limited beneath the deep snow pack, an important mechanism of ice growth during the N-ICE2015 winter was shown to be snow-ice formation 38,46,48 (Fig. 3). Snow-ice formation requires negative freeboard, caused by a heavy snow load that pushes the ice surface below sea level. In addition, a pathway is required for seawater to flood the ice surface (Fig. S3). Snow-ice formation is common for Antarctic sea ice where the seasonal ice cover is relatively thin and there is ample snowfall, but is not considered typical for the comparatively dry Arctic 49 . Observed snow-ice layer 48 Observed ice thickness 42 www.nature.com/scientificreports www.nature.com/scientificreports/ In the Atlantic Sector of the Arctic Ocean, frequent storms can contribute sufficient snowfall to induce negative freeboard. Rösel et al. 42 observed extensive areas of negative freeboard during N-ICE2015. Large areas with negative freeboard remained dry with no observed flooding 42 , but several autonomous buoys detected snow flooding shortly after strong sea ice deformation events during certain storms 38 (Fig. 3b,c). Seawater likely flooded the snow through cracks formed during sea ice convergence, triggered by storms. These flooded snow layers began to freeze and form snow-ice during the second phase of the storm, when cold and dry air was advected from the north 38 . Sea-ice cores from N-ICE2015 revealed saline layers of snow-ice of up to 0.3 m thick, demonstrating for the first time that snow-ice contributed significantly to the sea ice mass balance of Arctic pack ice 48 .
Here we explore the impact of changing winter precipitation on snow-ice formation, using our simulations with the 1-D sea ice model. When forced with the total precipitation from an atmospheric reanalysis, the model simulates a 0.24 m thick layer of snow-ice for the N-ICE2015 winter (Fig. 4). This is in agreement with ice core anlayses 48 . In contrast, when the total precipitation is reduced by 50%, the simulated snow load is insufficient to cause negative freeboard and no snow-ice forms in the model (Fig. 4). These simulations indicate that the potential snow-ice contribution (0.24 m) during N-ICE2015 could approximately compensate for the reduced thermodynamic growth from snow insulation (0.25 m), relative to a case with climatological snow depth. However, this model assumes that all areas with negative freeboard are flooded, which we know is not truly representative because large areas with negative freeboard were never flooded during N-ICE2015 42 . Hence, the total snow-ice contribution to ice mass balance is unlikely to reach its full potential and fully compensate the effects of snow insulation.
We note that the presence of saline snow and also the internal structure of the snow pack (e.g. ice crusts observed during N-ICE2015 43 ) can cause overestimation of sea ice thickness by space-borne radar altimeters such as CryoSat-2 30,50 . King et al. 30 showed that radar reflections from CryoSat-2 in the N-ICE2015 region during spring 2015 were closer to the snow freeboard than ice freeboard, and resulted in a systematic overestimation of sea-ice thickness by a factor of two. This highlights the need to give careful consideration of snow properties when interpreting satellite radar altimetry data in this region.
We speculate that continued thinning of Arctic sea ice may increase the potential for future flooding and snow-ice formation in the Atlantic Sector of the Arctic Ocean. Nonetheless, historic and projected trends in Arctic precipitation are highly uncertain 27,51 . With large regional differences across the Arctic, the net effect of thinning ice and changes in precipitation remain unknown. Later freeze-onsets and/or a trend towards more liquid precipitation (i.e. rain) could instead act to reduce the future snow load on sea ice and decrease the potential for snow-ice formation 14,27,52,53 . Winter storms open leads enabling new ice growth. Storm-driven ice deformation events are important for creating leads. During the first phase of each N-ICE2015 winter storm, southerly winds pushed the ice northwards into the ice pack 54 . This compacted the ice pack and increased the ice concentration around the N-ICE2015 camp (Figs 3a and 5a,b). This deformation was shown to predominantly occur along refrozen leads where the ice was weakest, but the compression and shear also cracked thicker snow-covered ice floes 55 . Due to this compression, the ratio of ice drift speed to wind speed was found to be lower during the first phase of each storm 54,55 .
During the second phase of each storm, northerly winds transported the sea ice southwards, towards the ice edge, promoting free drift 54,55 (Fig. 3a). Without resistance from the ice pack, the southward ice drift was found to be more rapid and generated strong divergence, opening new leads in the broken ice 54,55 (Fig. 6). These processes can be seen by the drop in sea ice concentration around the camp, during the second phase storms M1-M3, and M5-M6 (Fig. 5b). Similarly, following the second phase of storm M2 we see a large increase in the ratio of sea ice drift speeds to wind speeds 54 (Fig. 5a). The combination of open leads (i.e. lower ice concentration) and strong winds likely enabled large fluxes of heat, moisture, and gases between the ocean and atmosphere at these times 56,57 . For example, Fransson et al. 57 estimated that the CO 2 flux from the atmosphere into the ocean was approximately 20 times higher during storm periods, compared with average wind speed conditions and fewer open leads 57 .
Here we use high resolution satellite imagery to show the changing lead fraction in the vicinity of the N-ICE2015 camp, before (12%), during (9%), and after (28%) storm M2 (Methods, Fig. 6). Following the passage of each storm, any newly opened leads rapidly froze after cold and dry conditions returned, creating extensive areas of new ice. In-situ measurements from new ice forming in the leads opened by storm M2, revealed an ice thickness of 0.28 m after five days (Fig. 6c). We emphasise here that this new ice growth in leads would not have occurred without the aid of the storm, because growth rates under the existing snow-covered sea ice were negligible 38,42,46 (Fig. 3b,c). By assuming an average ice thicknesses of 0.28 m in newly opened leads, and 1.0-1.5 m for existing snow covered ice 42 , we estimate that storms M1 and M2 may have contributed each a 2-10% increase of sea ice volume in the vicinity of the N-ICE2015 camp (Fig. 6). However, the net effect of a storm on sea ice volume is complex, and will depend on local ice and ocean conditions as well as the path, size and strength of the storm.
Winter storms enhance ocean mixing, heat fluxes, and ice melt. Sea ice dampens energy transfers between the atmosphere and ocean, and therefore the Arctic Ocean is traditionally considered to be energetically 'quiet' with weak turbulent mixing 58 . However, strong winds during the N-ICE2015 winter storms enhanced ice drift speeds 54 (Figs 3a and 5a), which increased ocean-ice velocity shear 59 . These processes were found to generate mixing in the upper ocean, and led to increased transfer of ocean heat towards the ice 59,60 ( Fig. 5c-e). Observed winter ocean-ice heat fluxes typically more than tripled from 2 W m −2 to 7 W m −2 during storm periods (Fig. 5c, Methods), further impeding ice growth and in several cases initiating bottom melt 59,60 (Fig. 3b).
Ocean mixing is particularly important in many regions of the Arctic Ocean because warm water of Atlantic origin is located below cold and fresh Polar Surface Water 61 . Along the continental slope north of Svalbard, warm www.nature.com/scientificreports www.nature.com/scientificreports/ Atlantic Water (>2 °C) is found close to the surface (Figs 1, 5e). Vertical mixing thus generates enhanced ocean heat fluxes. The magnitude of this heat flux is dependent on the mixing rate, as well as the depth and temperature of the warm water. During the N-ICE2015 winter drift over the deep Nansen Basin, Modified Atlantic Water (>0 °C) was found at approximately 100 m depth 62 . Under calm conditions in the deep Nansen Basin, Meyer et al. 60 observed ocean heat fluxes at the pycnocline of approximately 3 W m −2 (Fig. S1). However, during storm periods, wind-driven mixing almost doubled the pycnocline heat fluxes to 5.5 W m −2 (Methods, Fig. S4a). These enhanced ocean heat fluxes are relatively small in comparison to changes in the atmospheric surface energy budget during storms 37 and were insufficient to induce ice bottom melt, but nevertheless acted to further suppress ice growth (Figs 2e and 5c). Previous work using autonomous buoy measurements have inferred enhanced ice-ocean heat fluxes during winter storms in the Beaufort Sea 25 . It is therefore expected that these conditions in the Nansen Basin are representative of large areas of the central Arctic Ocean.
During storm M3, the N-ICE2015 camp drifted over the Atlantic Water inflow path along the Svalbard continental shelf. Here, Peterson et al. 59 observed ocean-ice heat fluxes of 30 W m −2 , resulting in substantial ice bottom melt (Figs 3b and 5c-e). After the onset of bottom melt, the ice floe began to break-up and measurements were stopped to safely recover instruments. Larger heat fluxes would likely have been recorded if measurements had continued. Provost et al. 38 inferred ocean-ice heat fluxes from several buoys located near the N-ICE2015 camp at this time, which were on the order of 100-200 W m −2 (Fig. 3e). Similarly, during storm M6, a buoy located over the Atlantic Water inflow path experienced ice bottom melt rates of 25 cm day −1 , equivalent to an ocean-ice heat flux of 400 W m −2 (Fig. 3). An ocean mooring array in this region previously recorded episodic vertical ocean heat fluxes in excess of 100 W m −2 , during the winter of 2012-2013 63 , suggesting that ocean heat fluxes of this magnitude are not uncommon here. On both occasions that the N-ICE2015 camp drifted over the Atlantic Water inflow in winter, a storm was ongoing resulting in ocean-ice heat fluxes of 30 W m −2 due to the combined impact of storms with the presence of Atlantic Water at shallow depth (Fig. S4b). The timing of the storms also meant that winter-time ocean heat flux measurements over the Atlantic Water inflow are only available for stormy conditions, and not quiescent periods 59,60 . Due to strong tidal mixing in this region, we expect ocean heat fluxes over the Atlantic Water inflow path to be large even during quiescent conditions 19,64,65 . Nonetheless, spring and www.nature.com/scientificreports www.nature.com/scientificreports/ summer ocean heat flux measurements from N-ICE2015, and winter-time buoy measurements indicate a clear additional effect from storm driven mixing 38,59,60 .
We now investigate the large-scale impact of ocean-ice heat fluxes from storm-driven mixing over the Atlantic Water inflow path, using satellite observations of ice concentration and drift speeds (Methods). We find that each N-ICE2015 winter storm is associated with a reduction in sea-ice concentration of up to 25%, over the Atlantic Water inflow path north of Svalbard ( Fig. 7a-g). After ice advection is taken into account, we estimate that an ocean heat flux of up to 600 W m −2 would be required to fully melt the sea-ice area lost during storm M2 (Fig. 7g-i). During other storms, required heat fluxes range from 200-300 W m −2 . These values are of a similar order of magnitude as ocean-ice heat fluxes inferred from buoys located over the Atlantic Water inflow during N-ICE2015 winter storms 38 (Fig. 3e). Sea-ice concentrations typically return to pre-storm levels within a week of a storm (Fig. 7g,j), presumably after the reduction in wind speeds, ocean mixing, and ocean-ice heat fluxes allows ice growth to resume in open water areas and leads ( Fig. 5-8). However, the rapid succession of storms during the N-ICE2015 winter causes repeated ice loss in this region. Ivanov et al. 18 have previously shown that in winters with frequent and severe storms, like 2015, the sea ice concentration north of Svalbard remained low for several weeks, and in extreme cases open water extended along the path of the Atlantic Water inflow as far east as Franz-Josef Land. The N-ICE2015 observations demonstrate that storm driven ocean mixing provides an effective mechanism for releasing heat from the warm Atlantic Water inflow and melting sea ice during winter 59,60 .

Discussion
Legacy of winter storms for the arctic ecosystem. The spike in surface ocean CO 2 concentrations observed at the N-ICE2015 camp during and after storms M2 and M3, signals deeper CO 2 and nutrient rich water being mixed upwards 57 (Fig. 5d). Storm-driven ocean mixing during N-ICE2015 was found to enhance nitrate fluxes into the mixed layer, contributing to winter-restocking of nutrients for spring biological activity 66,67 .
The legacy of the six N-ICE2015 winter storms was an ice-pack formed of small floes covered by thick snow, and interconnected by numerous open or refrozen leads with very thin snow cover 68 (Figs 3 and 6). It is commonly understood that the spring algae bloom can only develop after the light-attenuating snow layer melts 69 . However, late spring observations from N-ICE2015, made in the same region but over different ice floes, showed that the increased light transmission through the numerous leads was sufficient for the development of an early under-ice phytoplankton bloom during the pre-melt season [70][71][72] (Fig. 8).
Pressure ridges created by deformation during winter storms were identified as another biological hotspot during the N-ICE2015 spring 73 (Fig. 8). The keels of ridges, especially those facing adjacent leads, provide a sheltered and favourable growth environment for ice algae 73 . Localised flooding caused by the heavy snow load and negative freeboard in spring was found to allow the infiltration of phytoplankton from the surface ocean to the snow-ice interface 73 (Fig. 8). These snow-infiltration communities are common in the Antarctic seasonal ice cover, but few reports exist in the Arctic 74,75 . As with snow-ice formation, we speculate that this phenomenon may become more common in the Atlantic Sector of the Arctic Ocean, if sea ice continues to thin and/or snowfall increases in the winter and spring 45,73,76 . www.nature.com/scientificreports www.nature.com/scientificreports/ storms precondition ice for faster summer melt. Our findings show that sea ice deformation and divergence during storms can promote new ice growth in leads, and may locally increase the sea ice volume. However, deformed ice is known to be more susceptible to melt 24 . Itkin et al. 54 recorded sea ice deformation rates around the N-ICE2015 camp that were 10% higher than those measured in other parts of the Arctic Ocean. Sea ice deformation breaks the ice into smaller floes that drift more easily with the wind 77 . Following storm M2, a 60% step increase in the ratio of ice drift speed to wind speed was observed 54 (Fig. 5a). The increased drift speeds, following storm driven sea ice deformation, will likely accelerate the transport of older and thicker sea ice out of the central Arctic and into the marginal ice zone of the Atlantic Water inflow path (Fig. 3a). In addition, smaller floe sizes following storms will precondition the ice for more lateral melt 24 . We further speculate that winter storms may contribute to an increased lead fraction in summer, reducing the albedo of the wider region. www.nature.com/scientificreports www.nature.com/scientificreports/ It is unclear how a deep winter snow pack, resulting from frequent winter storms, would ultimately influence the summer melt season. This could postpone the initial melt pond development at the snow-ice interface, resulting in a higher albedo and thus slowing the melt onset. Fresh snowfall from any spring storms would also increase the surface albedo and have a large influence on the melt onset 78 . After melt onset, we might expect a deep snow pack to result in a large melt pond fraction with low albedo, which would accelerate melt 79 . However, these melt ponds may drain rapidly, due to the heavily fractured ice floes. The N-ICE2015 campaign ended due to floe break up prior to the widespread development of melt ponds, and so it was not possible to study these processes. We highlight this as an important period to target for future field campaigns.

Conclusions
In this study we synthesise and build upon interdisciplinary results from N-ICE2015, to illustrate the multitude of ways in which winter storms accelerate the demise of sea ice in the Atlantic Sector of the Arctic Ocean (Fig. 8).
Winter storms transport pulses of heat and moisture into the Arctic atmosphere, which temporarily impede ice growth. However, ice-growth does not resume after these short-lived warming events. Snow deposited during storms insulates the ice from the cold atmosphere and reduces ice growth throughout the entire winter season. Strong winds break the ice into smaller floes that drift faster and will experience more lateral melt. Moreover, northerly winds, in the latter phase of storms, accelerate the transport of thicker-older ice in the Transpolar Drift towards the warm Atlantic Water inflow. Over the Atlantic Water inflow path, storm-driven mixing can generate large ocean heat fluxes that melt the ice from below, even in mid-winter. Hence, the cumulative effect of frequent winter storms in the Atlantic Sector of the Arctic Ocean is a thinner, weaker, and younger snow-laden ice-pack (Fig. 8).
The impact of an individual storm on sea ice is complex, and will be highly dependent on the timing, track, strength and lifespan of a storm. We conclude that storms which occur early in the winter (i.e. shortly after freeze-up) are likely to have a more negative impact on winter ice growth, compared with storms later in the season. This is due to a combination of the lasting effects of snow insulation, and the fact thinner sea ice is more sensitive to the effects of snow insulation than older thicker ice 29 Figure 8. Schematic of processes related to Arctic winter storms. In the first storm phase, strong southerly winds compress the ice cover and transport atmospheric heat, moisture, and precipitation into the Arctic. The atmospheric warming temporarily slows ice growth, and snow accumulation effectively inhibits bottom ice growth throughout the remaining winter season. If the snow load is sufficient to induce negative freeboard, snow may be flooded along the cracks in the ice during a storm. This flooded snow may then refreeze, forming snow-ice. In the second phase, northerly winds transport ice southwards towards an unconstrained ice edge. The rapid drift opens numerous leads, allowing intensive air-ocean gas and heat exchange and strong upperocean mixing, resulting in enhanced ocean heat fluxes. Along the Atlantic Water inflow path, storms mix up heat from warm subsurface waters and effectively melt large volumes of ice even in winter. After the storm has passed, cold and calm conditions return, allowing new ice to grow in leads. When the next winter storm arrives, it further drives the ice cover into a relatively thin-ice, snow-covered mosaic of strongly deformed ice floes that also impacts ice-associated ecosystems by shaping habitats and light conditions. www.nature.com/scientificreports www.nature.com/scientificreports/ the winter season, after a deep snow pack is established, may promote ice growth by dynamically opening leads where new ice growth can take place. Additionally, in cases when the snow depth is sufficient to induce negative freeboard, winter storms may trigger flooding and snow-ice formation.
While the observations presented in this paper are from storms in the Atlantic Sector, many processes we describe will be applicable for winter storms in other regions of the Arctic Ocean (Fig. 8). Specifically, the effects of snow-insulation, sea-ice deformation, and reduced radiative cooling, are likely to be important over the entire Arctic Ocean. We expect storm driven ice-ocean heat fluxes on the order of 5 W m −2 to be widespread over the Arctic Ocean 25 . These heat fluxes are unlikely to cause bottom melt, but will reduce winter ice growth. Large ocean heat fluxes capable of causing rapid bottom melt are a regional phenomenon unique to the Atlantic Water inflow path. Similarly, we only expect snow-ice formation to occur in the Atlantic Sector of the Arctic Ocean, where winter precipitation totals are sufficiently high 27 relative to the thinning sea ice to make it susceptible to flooding 15,42 .
The N-ICE2015 observations from the Atlantic Sector of the Arctic Ocean are especially important because this region experiences the most frequent and strongest winter storms 3 . While winter storm activity has increased here in recent years, it is unclear whether this trend will continue into the future 3,80 . Nevertheless, the overall impact of winter storms on sea ice will likely increase in a warmer Arctic. With a thinner sea-ice cover 76 and warmer and shallower Atlantic Water 19 , the tight air-snow-ice-ocean coupling we observed during storms will strengthen. Therefore, we expect more break-ups and ridging, more leads, greater ocean heat fluxes, faster melt, and possibly more flooding in response to winter storms. These emerging physical properties of the new Arctic sea ice regime will also have consequences for the ice-associated ecosystem.
The spatial scales of many processes highlighted above are typically small (<1000 m). This hampers the detection of these processes from remote sensing and ultimately their representation in climate models. In-situ observations are thus essential to improve our understanding of these processes. We emphasize that the autumn-freeze up and winter season are important periods to target in future field campaigns as it sets the scene for the summer melt period.

Methods
Overview of field campaign. The Norwegian young sea ice expedition (N-ICE2015) took place in the sea ice pack north of Svalbard from January -June 2015, onboard the Research Vessel Lance (Fig. 1). The ship was moored to an ice floe and a drifting ice camp was established at approximately 83°N. Due to rapid ice drift southwards to the marginal ice zone and floe breakups, four separate drifts were conducted. Here we analyse data from the first two drifts covering the periods 19 Jan -22 Feb and 24 Feb -20 March 2015. These distinct drift periods are highlighted in Figs 1-3, 5 and 7. For further details of the field campaign please refer to Granskog et al. 33 .
Meteorological observations. The N-ICE2015 meteorological observations 81 were first presented by Cohen et al. 34 . Wind speed and direction were measured on a 10 m mast at the ice camp, located approximately 300 m from the ship, using a 2D ultrasonic wind sensor (Lufft Ventus V200A-UMB). Mean sea level pressure was measured with an RM Young 61302 V pressure sensor at the base of the mast. The 2 m air temperature was measured with a Vaisala HMP155 sensor mounted inside a triple-walled actively aspirated radiation shield (RM Young Model 43502). We averaged these processed data from 10-minute resolution to six-hour means (Fig. 2). Precipitation was collected in a Tretyakov shielded gauge, mounted 1-1.5 m above the surface and located 10 to 15 m from the meteorological tower. The contents of the gauge were collected daily at ~0700 UTC, melted, and the water equivalent amount was measured to the nearest millimetre. Wind-under-catch loss is estimated to be on the order of 25 to 30% and can be up to 50% for high wind speeds 82 (>10 m s −1 ). Hence, the precipitation measurements are likely to be biased low due to blowing snow, especially during storm periods. For further details about these datasets, quality control, and uncertainties, refer to Cohen et al. 34 . For comparison with the precipitation observations, we also present the total daily precipitation from the European Centre for Medium Range Weather Forecasting's Interim Re-Analysis (ERA-Interim) 83 at the ice camp position. The timing of precipitation events in the Arctic is known to be very consistent among reanalyses, although there is a spread in the absolute values of total precipitation 27 . Note that we use the total precipitation product and not the snowfall. This is because a significant fraction of the total precipitation in the central Arctic is classified as rain, even during midwinter and sub-zero temperatures in ERA-I 27,46 . Downward longwave radiative fluxes 84 were first presented by Walden et al. 37 . These were measured using a heated and ventilated Kipp and Zonen CGR4 pyrgeometer 37 . We averaged the processed data files from 1-minute resolution to six-hour means (Fig. 2). The atmospheric moisture content was calculated by Kayser et al. 36 , using balloon borne data 85 . These radiosondes (Vaisala RS92) were launched from the ice camp daily at 1100 and 2300 UTC 36 .
Definition of a storm. The storm periods highlighted in Figs 1, 2, 3, 5 and 7 were first given by Cohen et al. 34 .
These storms (M1 -M6) were defined locally at the N-ICE2015 camp, as periods with a rapid pressure drop of at least 5 hPa/6 hrs and sustained wind speeds of more than 7 m s −1 (Fig. 2). In addition, we show the tracks of storms M1 -M6 in Figs S1-S2, classified by The University of Melbourne's cyclone identification scheme 35 using ERA-Interim reanalysis 83 .
Buoy observations. Several Sea Ice Mass Balance buoys (SIMBA), developed by the Scottish Association for Marine Research 86 , were deployed between 15-29 January as part of N-ICE2015 38,54,87 . The buoys were deployed on level ice and consisted of a 5 m long thermistor chain hanging through the air-snow-ice-ocean. The chain was comprised of solid state sensors that measured temperature with an accuracy of 0.1 °C at 2 cm vertical resolution every six hours. In addition, the sensors had a heating mode that recorded a proxy for thermal diffusivity. In this study we use data 88 from two buoys, SIMBA_2015a and SIMBA_2015f (Fig. 3). The interfaces 89 between ice-water, ice-snow and snow-air, the flooding of snow, and the ocean heat fluxes 89 were determined by Provost et al. 38 . We smooth all times series using a 36-h running mean. (2019) 9:9222 | https://doi.org/10.1038/s41598-019-45574-5 www.nature.com/scientificreports www.nature.com/scientificreports/ The ice drift speed in Fig. 3a was calculated using the GPS position of SIMBA_2015a 88 . Itkin et al. 54 calculated the mean ice drift speed in Fig. 5a using hourly GPS positions from 16 buoys (combination of SIMBAs and simple GPS drifters) deployed at locations within a radius of 120 km around the N-ICE2015 ice camp 88 .

Idealized model simulations.
We apply the 1-D thermodynamic sea ice model, HIGHTSI 90 , for a case study of N-ICE2015, covering the period September 2014 -March 2015. The design of the experiment is similar to the set up described by Merkouriadi et al. 46 . Where available, we force the model with N-ICE2015 observations. At other times, we use the ERA-Interim reanalysis 83 . An earlier comprehensive evaluation of ERA-Interim for the N-ICE2015 winter found good agreement between the reanalysis and observations 5 . For the period before the field campaign we apply the reanalysis along the satellite derived back trajectory 54 of the ice floe from Drift 1. We apply a constant ocean heat flux of 2.7 W m −2 throughout the simulation (a mean value suggested by McPhee et al. 91 ). We prescribe an initial ice thickness on 1 st September of 1.2 m, and snow depth of 0 cm. The initial ice temperature is assumed to decrease linearly from −0.25 °C at the ice top to −1.9 °C at the ice bottom. For further details of the experiment set up see Merkouriadi et al. 46 .
We conduct two idealised experiments to explore the impact of changes in precipitation on sea ice mass balance (Fig. 4). In the first simulation we use the total precipitation from ERA-Interim (1*Precip). In the next simulation we reduce precipitation by 50% (0.5*Precip). This simple case study only considers changes in the magnitude of precipitation. We do not explore how changes in the timing of storms (and thus precipitation events) during winter impact on ice growth. Further sensitivity experiments were conducted where the initial ice thickness was varied from 1.0-1.5 m and where the initial ice temperature profile was altered. We also explored different oceanic heat fluxes 46,91 . These changes did not alter any conclusions drawn. sea ice concentration at ice camp. Fransson et al. 57 derived the mean daily sea ice concentration around the ice camp from the AMSR2 microwave radiometer 92 (Fig. 5b). These data have a spatial resolution of 6.25 × 6.25 km 2 . The ice concentration was averaged over a 40 km square (7 × 7 grid cells), centred on the camp 57 .
Sea ice classifications from satellite remote sensing, sea ice thickness and new ice volume estimates. We use TerraSAR-X (TSX, © DLR 2015) synthetic aperture radar images to calculate the lead fraction around the N-ICE2015 camp. We process the images to 40 meters ground resolution and segment these using the method from Doulgeris and Eltoft 93 and Doulgeris 94 (Fig. 6). We subsequently classified the segments into two classes, using auxiliary information. The two classes are defined as leads (open water and thin ice) and the remaining ice cover (level and deformed ice).
We estimate the volume of new ice growth in leads opened by storms using the areal fraction of leads calculated in the segmented images, as well as ice thickness measurements taken within these leads on 12 and 14 February (on average 0.28 m thick). Lead ice thickness surveys were conducted by drilling. We assume an average sea ice thickness of existing snow-covered ice of 1.0-1.5 m, based on extensive ice thickness measurements from Rösel et al. 42 . Rösel et al. 42 measured the total thickness of snow and sea ice by electro-magnetic induction (EM31SH, Geonics Ltd.) on a large scale. From this, they calculated the ice thickness by subtracting the snow depth measured with a snow probe (Magnaprobe, Snow-Hydro) 42 . oceanographic observations. Ocean-to-ice heat fluxes 95 shown in Fig. 5 were measured with a Turbulent Instrument Cluster (TIC) installed at a depth of 1 m below the ice surface during Drift 1 and during a short period of Drift 2. The TIC was located a few hundred meters from the ship to avoid sampling in its wake. The TIC was equipped with a Sontek Acoustic Doppler Velocimeter (ADV), sampling at 24 Hz, to measure ocean currents. In addition, there were Sea-Bird Electronics sensors, SBE3F and SBE4, to measure temperature and conductivity at 24 Hz. Processing of the TIC data and calculation of ocean heat fluxes at the ice-ocean interface was performed by Peterson et al. 59 , and followed standard methods reported by McPhee et al. 91 .
We calculate the mean ocean-ice heat flux (from 1 m TIC) during the winter storms to be 6.7 W m −2 . Storm periods are based on Cohen et al. 34 , and are highlighted in Fig. 5. The mean ocean-ice heat flux (from 1 m TIC) during the winter period when there is no forcing (i.e. no storms, no steep topography below, and no shallow Atlantic Water) is 2.2 W m −2 . Mixing in the ocean is often still enhanced a few days after a storm (Fig. 5). Therefore, we define 'no storm' periods as the time between 3 days after the end of a storm and the day before the start of the next storm, to avoid any contamination.
Dissipation rates, mixed layer depth, and ocean heat fluxes at the pycnocline depth for the N-ICE2015 winter, shown in Fig. 5, were calculated by Meyer et al. 60 using microstructure profile observations 96 . Typically, 3 sets of 3 microstructure profiles were conducted daily between 0800 UTC and 2000 UTC, using two loosely tethered free-fall MSS-90 microstructure profilers 97 developed by ISW Wassermesstechnik. Each final set of microstructure profiles is therefore the result of 3 combined individual microstructure profiles. These sets were used to estimate the profiles of ocean heat fluxes presented in Fig. S4. The MSS-90 microstructure profilers contained precision conductivity, temperature, pressure, and microstructure sensors and sampled at 1024 Hz. Profiles were made through a hole in the ice adjacent to the TIC installation. Data processing followed the method of Fer et al. 98 . The processed data provide 0.2 m averaged temperature and salinity measurements and 1 m averaged dissipation rates. The mixed layer depth shown in Fig. 5e was calculated by Meyer et al. 62 .
The fCO2 observations 99 shown in Fig. 5 were processed by Fransson et al. 57 . The data were obtained by infrared analysis of equilibrator headspace samples. The seawater was supplied from an intake located midships, at approximately 5 m water depth. The measurements were made using an instrument supplied by General Oceanics and designed following the principles presented by Pierrot et al. 100 using two-stage showerhead equilibration and a LICOR 7000 nondispersive infrared detector 57 . (2019) 9:9222 | https://doi.org/10.1038/s41598-019-45574-5 www.nature.com/scientificreports www.nature.com/scientificreports/ Large-scale impact of storms. To investigate the large scale impact of the six N-ICE2015 winter storms, we use satellite sea ice concentration and drift vectors from the Ocean and Sea Ice Satellite Application Facility (OSI-SAF) and the ERA-Interim atmospheric reanalysis 83 . Specifically, we show the sea ice extent before and after each of the six N-ICE2015 storms. Together with this we show the 10 m wind speed and mean sea level pressure field averaged over the duration of each storm (Fig. 7). The direction of surface winds is typically aligned parallel to the mean sea level pressure isobars, and the flow is anti-clockwise around the low pressure centre, marked with an (L).
Using the OSI-SAF sea ice drift product with a 62.5 km resolution, we calculate the import/export of sea ice area into a box covering the Atlantic Water inflow region north of Svalbard, throughout the N-ICE2015 winter (Fig. 7). We define our Atlantic Water box on the OSI-SAF drift vector grid. The four corner coordinates are: 5.19°E, 82.80°N; 18.43°E, 79.71°N; 39.29°E, 80.74°N; 35.54°E, 84.39°N. To calculate the net ice area import or export within the box, we multiply the ice drift velocity from OSI-SAF perpendicular to the box edge with the sea ice concentration at that grid point and the grid length. We sum this total around the perimeter of the box. Using the OSI-SAF sea ice concentration product, we calculate the daily change in sea ice area within the box. Note, the time period for averaging over the storms in Fig. 7a-f is chosen to coincide with the reduction in mean sea ice concentration within this box (Fig. 7g). From the daily change in ice concentration, we subtract the area of ice import/export to the box. Finally, we make the assumption that the remaining area loss of sea ice is caused by ice melt, due to storm driven mixing of warm Atlantic Water. We calculate the required ocean heat flux to melt this ice, assuming three different constant ice thicknesses to the sea ice area (1.00 m, 1.25 m and 1.50 m). We use a constant ice density of 900 kg m −3 , and use the specific latent heat of fusion of pure ice of 333500 J kg −1 multiplied by 0.89, following Bitz and Lipscomb 101 .
Caution should be taken with these heat flux estimates, because we do not account for ice lost due to convergence (e.g. ridging or rafting). However, we expect this term to be small compared to the ice-melt term. For comparison, ice export from the box contributed a maximum of 25% of the total area loss. Our method also does not account the melt of ice floes that only partially melt from below (e.g. Fig. 3b), since this is not detected in ice concentration measurements. The uncertainty of the ice concentration data for low ice concentrations in the marginal ice zone can exceed 10% 102 ; this adds further uncertainty to the ice-melt estimate.