To determine the origin, pathways and thickness changes of sea ice, as well as the atmospheric forcing acting on the ice cover, we use our Lagrangian drift analysis system called IceTrack that traces sea ice backward in time using a combination of satellite-derived, low-resolution drift products (Krumpen et al., 2019). The approach has also been applied in a number of previous studies for the same purpose (Ricker et al., 2018; Damm et al., 2018; Peeken et al., 2018; Krumpen et al., 2016 and others). In summary, IceTrack uses a combination of three different publicly available ice drift products for the tracking: (i) motion estimates based on a combination of scatterometer and radiometer data provided by the Center for Satellite Exploitation and Research (CERSAT; Girard-Ardhuin and Ezraty, 2012), (ii) the OSI-405-c motion product from the Ocean and Sea Ice Satellite Application Facility (OSI SAF; Lavergne, 2016), and (iii) Polar Pathfinder Daily Motion Vectors (v.4) from the National Snow and Ice Data Center (NSIDC; Tschudi et al., 2016). The contributions of individual products to the used motion field are weighted based on their accuracies and availability which vary with seasons, years and study region. The IceTrack algorithm first checks for the availability of CERSAT motion data within a predefined search range. CERSAT provides the most consistent time series of motion vectors starting from 1991 to present and has shown good performance on the Siberian shelves (Rozman et al., 2011). During summer months (June–August) when drift estimates from CERSAT are missing, motion information is bridged with OSI SAF (2012 to present). Prior to 2012, or if no valid OSI SAF motion vector is available within the search range, NSIDC data are applied. The tracking approach works as follows: ice in user-defined individual starting locations or positions on a 25 km EASE2 grid is traced backward in time on a daily basis. Tracking is discontinued if (a) the tracked ice reaches the coastline or fast ice edge or (b) the ice concentration at a specific location along the backward trajectory drops below 40 % and we assume the ice to be formed.

To investigate the impact of winter sea ice dynamics on the summer ice cover, we calculate monthly sea ice area fluxes through the northern boundary of the Laptev Sea for the winter season from March to April (1992–2019). The gate is located between 110 and 160∘ E at 77.5∘ N (black line with arrows in Fig. 2a). The flux calculations follow the approach of Ricker et al. (2018), who estimated volume fluxes through the Fram Strait. For ice concentration, we use the CERSAT product. For ice motion, we use merged products from CERSAT that are based on radiometer and scatterometer data. Figure 2c shows the total ice area export from March to April of each winter, including a trend line plotted on top.

The timing of sea ice break-up and freeze-up (Fig. 2b) was estimated for each year based on CERSAT sea ice concentration data for the region between 86∘ N, 100∘ E and 71∘ N, 160∘ E. An ice-free grid point is defined as the first day in a series of at least 10 d when ice concentration exceeds and reaches zero (Janout et al., 2016).

Sea ice retreat during the melting period in the Laptev Sea and East Siberian Sea is the result of atmospheric and oceanic processes and regional feedback mechanisms acting on the ice cover, in both winter and summer. In the following, we will briefly review the sea ice conditions on the Siberian Shelf seas prior to the start of the expedition and the main preconditioning mechanisms that contributed to the northward retreat of the ice edge in 2019. In this regard, our focus is on the atmospherically driven processes, since results from oceanographic surveys are not yet available.

Ice dynamics and ice export in winter are important preconditioning mechanisms for the ice retreat in summer. Itkin and Krumpen (2017) observed that enhanced offshore-directed transport of sea ice in late winter has a thinning effect on the ice cover. During late winter months dominated by an offshore-directed drift component, newly formed ice areas are larger and remain comparatively thin and therefore melt more rapidly once temperatures rise above freezing. This feedback mechanism is even more pronounced when temperatures at the end of winter are unusually high. Figure 2 summarises the conditions and processes that shaped ice formation in the Laptev Sea and East Siberian Sea in winter 2018/2019. Satellite-based estimates of offshore-directed sea ice area transport between March and April are shown in Fig. 2c (1992–2019, from 110 to 160∘ E at 77.5∘ N). Late winter flux estimates indicate that the sea ice advection away from the Siberian shelves towards the central Arctic was approximately 70 % higher (2.32×105 km2) in 2019 than the long-term mean annual rate (∼1.36×105 km2). Following Krumpen et al. (2013), the strong positive trend (+0.53×105 km2 per decade) in late winter ice area export is associated with an increasing drift speed as a result of thinning ice cover and a rapid loss of thick multi-year ice. As a consequence of the intensified ice advection shortly before spring break, satellite-based sea ice thickness observations (Fig. 2a) show negative thickness anomalies throughout the entire coastal zones of the East Siberian Sea and the Laptev Sea in April 2019, except for the southern half of the area around the New Siberian Islands.

Ocean-driven preconditioning mechanisms are less well understood. However, there is indication that enhanced winter ventilation of the ocean can reduce sea ice formation in this area at a rate now comparable to losses from atmospheric thermodynamic forcing (Polyakov et al., 2017). Observations carried out in the eastern Eurasian Basin have shown that weakening of the halocline and shoaling of intermediate-depth Atlantic water layer result in heat flux equivalent to 40–54 cm reductions in ice growth in 2013/2014 and 2014/2015.

In addition, anomalously high temperatures during the winter months can further reduce the growth of first-year ice (FYI), resulting in thinner ice cover at the end of the winter (Ricker et al., 2017). According to NCEP reanalysis data (Fig. S1, Supplement) and observations from the Kotelny meteorological station (Fig. 2a, yellow circle), the temperatures during the ice growth phase (October 2018–May 2019) were elevated: reanalysis data show positive temperature anomalies of 3 ∘C in comparison to the 1981–2010 climatology, and records at Kotelny show significantly higher temperatures than those at the beginning of the instrumental record (Fig. 2e). In particular, temperatures at the end of the winter are unusually high. If this coincides, as described above, with periods of strong offshore-directed winds, the formation of new ice in coastal areas is reduced, which favours early melting of the ice cover in spring (Fig. S2, Supplement).

The subsequent temperature anomalies in spring and summer 2019 were even more pronounced. During the summer months, Kotelny meteorological monitoring station recorded the highest mean temperatures since the beginning of record-keeping (Fig. 2d), and the reanalysis data indicate a positive anomaly of 2.5∘ on the Siberian shelves and in adjacent northern regions (Fig. S1, Supplement). The rapidly rising temperatures in spring accelerated the melting of the ice cover, which was extremely thin to begin with (Fig. 2a). This resulted in the earliest ice break-up ever observed (compare Fig. 2b, red line) and rapid northward retreat of the ice edge, which exposed surface waters to direct solar heating. Consequently, summer (August 2019) sea surface temperatures south of the MOSAiC starting area were approximately 2–4 ∘C higher than the 1982–2010 mean (Timmermans and Ladd, 2019), such that wind events that force ice floes back into warm waters could have caused additional ice melt (Steele and Ermold, 2015). Moreover, the intensive warming of the upper ocean (Janout et al., 2016) caused a delay in the autumnal freeze-up of sea ice (Fig. 2b, blue line) and resulted in large parts of the marginal seas remaining ice-free for up to 93 d. This means that the MOSAiC expedition started immediately after the longest recorded ice-free period in the region.

In this section we describe the predominant ice conditions at the beginning of MOSAiC, in both the ship's immediate vicinity and its extended surroundings. The latter encompass the area within a 220 km radius of Polarstern and will hereafter be referred to as the extended MOSAiC region (EMR; see Fig. 3a). A radius was selected to include various ice types, which differ in terms of their provenance (i.e. origin) and/or age. The EMR includes both the ice edge to the south and thicker and more stable pack ice to the north. The ship's immediate vicinity (distributed network region, DNR) includes the DN and has a radius of 40 km. We will first describe the ice conditions in the EMR, before turning our attention to the DNR.

Once the MOSAiC floe had been chosen, we applied a tracking tool (see Methods) to the residual ice that was in the EMR shortly before MOSAiC's starting date. Figure 3b shows the age of the sea ice within the EMR on 25 September. Based on the backtracking analysis, the EMR's residual ice had an average age of 318 d and was formed on 11 November 2018 (±15 d). Second-year (SYI) or multi-year ice (MYI) was not found, either from tracking or from scatterometer data. Most of the residual ice was originally produced during or shortly after the freeze-up in polynyas (or elsewhere on the shallow Siberian shelves) (Fig. 3c), featuring water depths of less than 30 m. Only the ice at the far eastern and northern edges of the EMR originated from regions with a water depth exceeding 50 m. From the time of its formation to 25 September, the EMR ice had travelled an average distance of approximately 2440 km (±205 km, Fig. 3d) and experienced low ice concentrations between June and September 2019 (Fig. 3e). Hence, the residual ice encountered after our arrival on site was severely weathered, and bridge observations indicated that a large fraction was melted completely during summer months. Residual ice that survived was characterised by frozen-over melt ponds with a <10 cm deep layer of fresh snow. Based on visual observation, melt pond fraction was 70 %–80 % in the undeformed ice areas, and the bottom layer experienced internal melting. According to ice coring, only the top 30 cm of ice was solid. Because both ships only reached the target region after the freeze-up had begun, large expanses of previously open water were now covered with new ice.

Based on the backtracking analysis, the floes selected for the Central Observatory and the DN were located in a zone of comparatively young ice that formed roughly 3 weeks later than the ice within the EMR (Fig. 3b, early December 2018) and originated from a shallow (Fig. 3c) region closer to its location on 25 September (Fig. 3d, 2240 km). Figure 4a shows the trajectories obtained for the centre of the DNR (the position of the CO, red line) and four adjacent positions at a distance of 25 km (grey lines). Information on water depths and ice concentration along the central trajectory is provided in Fig. 4b, c. The trajectories indicate that the ice inside the DNR was formed in a polynya event on 5 December 2018, north of the New Siberian Islands in water that was less than 10 m deep. An eastward ice drift then transported the newly formed ice along the shallow shelf, until it reached deeper water in early February 2019. Ice cores collected at various points in the DN and on the CO confirm that the DNR ice originated in the shallow Siberian shelves, since some of the cores contained sediment inclusions of sandy silt in the uppermost 50 cm (Fig. 4c, d). Though the quantities were small in most cases, these inclusions can only be found on the shallow Siberian Arctic shelves with average water depths of less than 30 m (Sherwood, 2000; Wegner et al., 2017). There, particulate matter and organisms are incorporated into the newly formed ice by suspension freezing (Eicken et al., 2000) or, to a smaller degree, by grounded sea ice pressure ridges ploughing through the sea floor (Darby et al., 2011). A detailed chemical analysis of these trapped sediments will be conducted at a later point in time.

The validity and reliability of Lagrangian drift studies depend on the accuracy of the applied sea ice motion product. In this study, we primarily use the CERSAT drift dataset because it provides the most consistent time series of motion vectors starting from 1991 to present (see Methods). Comparisons with buoys and high-resolution SAR images indicate that in particular during winter months, when the atmospheric moisture content is low and surface melt processes are absent, the quality of motion products from low-resolution satellites is high (Sumata et al., 2014; Krumpen et al., 2019). Restrictions may arise from the coarse resolution of the sensors in near-shore regions characterised by a complex coastline, extensive fast-ice areas, and polynyas (Rozman et al., 2011). During summer months (June–August), when strong surface melt processes and high moisture content in the atmosphere further reduce accuracy of low-resolution motion products (Sumata et al., 2014), IceTrack uses the OSI SAF motion product to bridge the lack of CERSAT data. To quantify uncertainties of sea ice trajectories on a larger temporal and spatial scale, we reconstructed the pathways of drifting buoys using IceTrack. For this purpose, we selected 10 buoys that had survived a full summer and winter in the Arctic. Their drift was then reproduced from October onwards in a backward direction over 12 months. Figure S3 (Supplement) shows the deviation between actual and virtual tracks, which is rather small (60±24 km after 320 d) and in an acceptable range. The maximum deviation between real and virtual buoys gives a measure of the largest possible error that can occur when determining the ice origin. After 320 d it is around 105 km. The confidence bound is shown in Fig. 4 as an ellipsoid (dashed line). No significant differences in sea ice pathways and source areas were observed when repeating the tracking experiment using different combinations of motion products, or higher and lower ice concentration thresholds.

Note that we originally planned to trace the provenance of the MOSAiC floe using high-resolution satellite data (Sentinel-1, TerraSAR-X and MODIS). However, only sporadic high-resolution images of the region were available, and the combination of low summertime sea ice concentration and high degree of cloud cover made it extremely difficult to manually track the exact position of individual floes over an extended period of time. Nevertheless, the high-resolution satellite data enabled us to track nearby large-scale patterns such as shear zones or very prominent floes. Hence, we could at least determine the approximate location of the MOSAiC floe on individual images. The resulting estimates for the different positions of the CO (brown-yellow coloured circles in Fig. 4a) correspond well to the computed trajectories (red line in Fig. 4a), which lends increased confidence in our results.

To calculate the ice thickness variability in the EMR and DNR at the start of MOSAiC (Fig. 5a) and the ice thickness evolution along the drift trajectories encountered by the ice in those regions (Fig. 5b), we used the results of a thermodynamic model (see Methods). Results show that the residual ice in the DNR was not only younger and originated from a different location than the ice in the surrounding EMR, but it was also thinner: on 25 September, the averaged ice thickness inside the EMR was 0.58 m (±0.27 m), while the thickness of ice inside the DNR was 0.37 (±0.09 m), i.e. 36 % (0.21 m) less than in the EMR. To confirm model results, we applied a second, simpler thermodynamic model developed by Thorndike (1992) and used in Peeken et al. (2018) and Krumpen et al. (2019). The model is chiefly based on air temperatures, assumes a constant ocean heat flux and employs snow climatology, but indicates the existence of similar thickness gradients between the EMR and DNR (40 % difference; results not shown here). Nevertheless, the decrease in ice thickness toward the DNR is clearly recognisable in both models and is in agreement with direct field observations: Fig. 6 shows the results of the GEM ice thickness measurements carried out on four floes in the distributed network (L1–L3 and M8) and compares them with measurements taken on R1. The measured difference in modal ice thicknesses (without snow) between R1 and the DNR was 0.3 m (R1: 0.5 m vs. DNR: 0.2 m). Higher ice thicknesses were also measured at R2 and R3 located farther to the north and west, which were reached by helicopter (Table 1, Methods).

Visual observations made from the bridge of the Akademik Fedorov as it travelled along the expedition route provided further evidence for the presence of a thickness gradient between the DNR and EMR. The percentage of residual ice steadily dropped from nearly 90 % at R1 to 20 % at the DNR; conversely, the percentage of thin, newly formed ice rose from 10 % to ca. 80 %. This indicates that, given its lower initial thickness at the end of the winter, some of the ice in the DNR could have completely melted in summer. The thickness gradient between the DNR and EMR is confirmed by CryoSat-2–SMOS measurements from the end of winter 2018/2019. Already in April 2019, a negative thickness anomaly prevails at the later starting position of the drift experiment (Figs. 2a and 3f).

We showed that due to its younger age and different provenance, the DNR ice was thinner than the surrounding ice. But the thicknesses measurements summarised in Fig. 6 and Table 1 are also much smaller than what was observed by Haas and Eicken (2001) in the 1990s by similar GEM and drill-hole measurements. They found late-summer modal FYI thicknesses between 1.25 m (1995), 1.75 m (1993), and 1.85 m (1996) in regions near or south of the MOSAiC study region, supporting the notion of exceptionally thin ice in the MOSAiC starting region. In this section, we compare the conditions we encountered at the end of September 2019 with those of previous years by applying the combined tracking–thermodynamics model to the period between 1994 and 2019. Figure 7a shows the history and variation in imaginary MOSAiC floe trajectories for the past 26 years. Tracking was performed backwards in time starting from the DNR region on 25 September of each year. Results indicate that the climatological probability that DNR ice originates from the New Siberian Islands, like in 2019, is about 25 % (red shaded area and tracks). From a climatological perspective, it is usually more likely that the ice at the starting position has its origin in the Laptev Sea (55 %, light blue shaded area). A smaller part (∼20 %) typically comes from the East Siberian Sea (grey shaded area). The approximate age of the ice near the starting point is around either 1 or 2 years (Fig. 7b), with a tendency towards decreasing ice age. This tendency of decreasing ice age is evident from the frequency of SYI. While SYI occurred in about 64 % of all years between 1992 and 2004, it was already much less frequent during the past 15 years (20 %, 2005–2019).

Figure 7c displays the time series of September FYI thickness estimates in the DNR for the period between 1994 and 2019. In addition, Fig. 7d provides the annual cycle of DNR ice growth and melt. An overall decrease in residual ice thickness between 1994 and 2019 is visible (trend: -0.22 m per decade), which is subject to a high interannual variability and therefore not statistically significant. The DNR ice encountered in September 2019 can be classified as exceptionally thin over a longer period of time (Fig. 7c). However, for the larger region of the EMR, ice thicknesses in September 2019 agree well with the long-term average (Fig. 7d). Both DNR and EMR ice shows above-average growth rates in winter 2018/2019 as well as above-average thicknesses at the end of April, followed by above-average melt. An in-depth analysis of the applied forcing data in the thermodynamic model reveals that the intensified ice production is a consequence of reduced precipitation rates in winter 2018/2019 (Fig. S4, Supplement).

Through a comparison with in situ data, we have shown above that the thermodynamic model is able to simulate regional differences in ice thickness. However, in order to verify that the model is capable to reproduce the interannual variability correctly, model estimates require comparison to historical observational data from the past. Unfortunately, field surveys in this exact location and that time of the year are scarce, but GEM ice thickness measurements in the surroundings of the DNR between 84 and 86.5∘ N and 100 and 150∘ E (compare Fig. 3) were obtained by Haas and Eicken (2001) during the ARK-12 cruise of Polarstern in August 1996. The authors obtained around 37 km of thickness profile data at 5 m horizontal spacing. They found average FYI modal thicknesses of ∼1.85 m, typical for SYI or even MYI in summer. The 1996 GEM measurements were obtained 6 weeks earlier in the melt season (10 to 22 August 1996) inside the EMR area and south of it. In comparison, the exceptionally thick September 1996 ice is reproduced by our thermodynamic model with 1.6 m in the DNR (Fig. 7c). According to Haas and Eicken (2001), the relatively thick ice in 1996 was due to specific atmospheric circulation conditions during summer, characterised by persistent low sea level pressure over the central Arctic. This resulted in very weak surface melt and the absence of melt ponds north of approximately 84∘ N in 1996. The model results and forcing data for 1996 confirm that strongly reduced net shortwave fluxes led to a significant reduction in ice melting during the summer months. Even in years dominated by strong melting processes, the model seems to realistically reproduce ice thickness: in winter 2013/2014, ice formed comparatively late in the season and melted completely during summer (Fig. 7d). Satellite sea ice concentration data confirm that the DNR region and large parts of the EMR were ice-free already at the beginning of August 2014. If combined with reliable trajectory and realistic forcing data, the good agreement between the thermodynamic model and observations for the years 1996, 2014 and 2019 shows that the model can be used to study interannual variability of FYI thickness changes and the driving mechanisms behind them.