Carbon dioxide sink in the Arctic Ocean from cross-shelf transport of dense Barents Sea water

Large amounts of atmospheric carbon can be exported and retained in the deep sea on millennial time scales, buffering global warming. However, while the Barents Sea is one of the most biologically productive areas of the Arctic Ocean, carbon retention times were thought to be short. Here we present observations, complemented by numerical model simulations

Large amounts of atmospheric carbon can be exported and retained in the deep sea on millennial time scales, buffering global warming. However, while the Barents Sea is one of the most biologically productive areas of the Arctic Ocean, carbon retention times were thought to be short. Here we present observations, complemented by numerical model simulations, that revealed a deep and widespread lateral injection of approximately 2.33 kt C d −1 from the Barents Sea shelf to some 1,200 m of the Nansen Basin, driven by Barents Sea Bottom Water transport. With increasing distance from the outflow region, the plume expanded and penetrated into even deeper waters and the sediment. The seasonally fluctuating but continuous injection increases the carbon sequestration of the Barents Sea by 1/3 and feeds the deep sea community of the Nansen Basin. Our findings combined with those from other outflow regions of carbon-rich polar dense waters highlight the importance of lateral injection as a global carbon sink. Resolving uncertainties around negative feedbacks of global warming due to sea ice decline will necessitate observation of changes in bottom water formation and biological productivity at a resolution high enough to quantify future deep carbon injection.
Without the buffering function of the world oceans, the impact of anthropogenic greenhouse gas emissions would be far more dramatic 1 . The retention time of up to 25% of human CO 2 emissions from the atmosphere 2,3 strongly depends on the local export efficiency and sequestration depth. Photosynthetically fixed particulate organic carbon (POC) can increase retention times drastically due to gravitational sinking into deeper waters or sediments isolating it from the atmosphere for long time scales 4 . POC fluxes are traditionally considered to decrease exponentially with depth due to ongoing grazing and respiration 5 . However, a major uncertainty in such vertical Article https://doi.org/10.1038/s41561-022-01069-z (UVP) 17 camera system to measure particle size distribution and abundance. Our data showed a widespread plume of particles in the Nansen Basin, which was most pronounced north of Severnaya Zemlya (Fig. 1), where the maximum volume of particles 0.102-2.05 mm equivalent spherical diameter (ESD; further particles 0.1-2 mm) reached 3.5 mm³ l −1 (ppm). This plume extended over a cross section of 19.5 km 2 between 500 m and 1,200 m and more than 40 km off the shelf and was co-located with Barents Sea Bottom Water (BSBW) characterized by absolute salinities <35.06 g kg -1 and potential density anomalies >27.97 kg m −3 (Fig. 1c).
BSBW is formed by the transformation of North Atlantic waters during its passage through the Barents Sea due to cooling and mixing with shelf waters and brine, rejected during sea ice formation 10,13 . Water mass definition was based on this transformation: after leaving the shelf system BSBW north of Severnaya Zemlya was denser than waters in the intermediate density layer and less saline than dense Atlantic water from Fram Strait (Methods).
After leaving the shelf through St. Anna Trough, BSBW integrates into the respective density horizon below the Atlantic core of the Arctic Boundary Current. Steered by the Coriolis force, it propagates along the shelf break further east towards our study area 11,12 . The particle concentrations were low in depths shallower than 500 m, typical for Atlantic water in this region (Fig. 1b). Increased concentrations of particles below were hence transported laterally with the advection of BSBW. The presence of particles >0.1 mm, mainly in deeper layers of BSBW or below (Fig. 1b, Extended Data Fig. 1b-d) indicated gravitational sinking, resulting in a spatial separation during water mass propagation from the outflow region located several tens of km upstream. Increased turbidity exhibited a higher co-location with BSBW signals (Fig. 2), pointing to a high load of small particles in suspension within this water decay models is lateral transport, which, if substantial, would cause an increased input of POC at depth and decouple sources and sinks. These processes are rarely studied and generally not considered in global biogeochemical models. In the Amerasian basin of the Arctic Ocean, eddy transport 6 and advection of benthic nepheloid layers 7,8 have been identified as lateral transport processes to moderate depths, which might also partly explain a reported mismatch between local primary production and benthic carbon demand in the Arctic deep sea 9 . In the Eurasian basin, however, comparable studies are missing, although dense bottom waters, produced on the Russian shelf 10,11 , have the potential to inject an important carbon signature directly into the deep sea due to density-driven subduction. Notably, the Barents Sea is a hotspot of dense water formation 12,13 but also CO 2 uptake and fixation. With 36.4 Mt C yr −1 , it contributes ~20% of the pan-Arctic CO 2 uptake (180 Mt C yr −1 ) (ref. 14 ) and with 102 Mt C yr −1 (ref. 15 ), ~50 % of the total primary production 16 , but due to the lack of data from the dense water outflow region, its impact on the Arctic carbon budget is still not resolved. Here we present results of the campaign ARCTIC2018 onboard RV Akademik Tryoshnikov in the Nansen Basin and the Laptev Sea. Combining underwater imaging with water mass and circulation measurements, molecular sequencing techniques and model calculations, we show that dense water-driven carbon injection may provide an important pathway for Arctic carbon sequestration.

Bottom water-associated lateral particle transport
In late summer 2018, we performed transects of vertical profiles within the Nansen Basin east of St. Anna Trough (Fig. 1a)

Fig. 1 | Deep particle plume and its relation to BSBW north of Severnaya
Zemlya. a, Map of the research area including CTD stations (red dots) with the shown transect closest to the outflow area (red rectangle). b, Spatial particle volume distribution of the size fraction 0.1-2 mm measured along the transect. The area above the solid (absolute salinity <35.06 g kg −1 ) and below the dashed isoline (potential density anomaly >27.97 kg m −3 ) indicates BSBW. c, Conservative temperature and absolute salinity plot with different colours showing particle volume of the size fraction 0.1-2 mm and relation of particle maximum with BSBW. Isolines indicate potential density anomaly (σ0) in kg m −3 . BSBW is located within red dashed water mass polygon. Particle volumes have the same colour bar in b and c.

Origin and evolution of the particle plume
To identify the origin of the material observed within the plume, we performed particle backtracking calculations using the Finite-Element Sea ice-Ocean Model (FESOM 18 ; Extended Data Fig. 2a-c), which confirmed a lateral injection from the shelf (Fig. 2c,d and Extended Data Fig. 2d-g). Reverse trajectories showed that large fractions of particles originated in the St. Anna Trough region on the shelf following the main ocean current field into the basin. More than half of the non-ballasted particle fraction <0.1 mm originated ~750 km upstream within the northeastern Barents Sea close to Novaya Zemlya (Extended Data Table 1). Open waters of lee polynyas with increased sea ice formation rates in this area are a known hotspot of BSBW formation 12,13 . Furthermore, enhanced primary productivity compared with the Arctic basin due to increased light availability within the marginal ice zone and polynyas in combination with nutrients from Atlantic waters, freshwater run-off and wind-induced vertical mixing 19-21 is characteristic for this region. On the other hand, deep mixing together with enhanced aggregation leads to high export efficiencies 22-25 , whereas strong mean and tidal currents keep the material in suspension in the frictional bottom boundary layer, which prevents burial in the sediment 26 and promotes resuspension.
The FESOM simulations included a warm bias of ~1 °C, leading to an underestimation of BSBW production and therefore also of the shelf-originating particle load. However, several lines of field evidence support the non-local origin of the particles: (1) 16S and 18S rRNA gene amplicon sequencing revealed that the microbial community of the suspended particle fraction <0.1 mm differed significantly in the plume (P < 0.01; Extended Data Fig. 3). Moreover, (2) we identified the benthic indicator species Dolichomastigaceae, recently isolated from deep sea sediments 27 , to be associated with those particles (Supplementary Table 1). (3) Increased fluorescence signals in the plume (Fig. 2b) point towards a phytoplanktonic origin of fractions of the particle load. (4) The majority of marine snow >0.8 mm within the plume consisted of degraded phytodetritus aggregates as identified in UVP images (Extended Data Fig. 1). Their distinct circular and compact morphology has been described as characteristic for processed aggregates during ending Arctic phytoplankton blooms 28 . (5) Abundances of copepods >0.8 mm were increased within the plume (Extended Data Fig. 1g). UVP images showed intact and healthy-looking copepods, suggesting they were alive and probably actively grazing. Synthesizing this list of field evidence with our model results above, the particle load transported with BSBW can be described as a mixture of fresh material from the phytoplankton-grazer community within the euphotic shelf zone and resuspended material from sediments.
As a result of this mixed pool of non-local relatively fresh and old material, also particulate C:N ratios revealed contradictory results: low C:N ratios of particles <0.1 mm are typical for cold, nutrient-rich, high-latitude waters 29 , possibly promoted by diatoms 30 and heterotrophic eukaryotes 31 as identified within our amplicon sequences (Supplementary Table 1). Particles >0.1 mm, on the other hand, showed C:N ratios up to >40 (Extended Data Fig. 4), especially within the transect closest to the BSBW outflow region. Because surface samples-but not the fraction <0.1 mm-also were affected, we suggest extrapolymeric substances, such as polysaccharides within large phytoaggregates 32 , to have caused these values. These substances are carbon-rich, nitrogen-poor and buoyant, which might also support long-range lateral transport due to reduced sinking velocities.    In the course of the campaign, we could follow the path of BSBW and its particle load further along the shelf break. Approximately 250 km downstream the first observation, east of Severnaya Zemlya (Extended Data Fig. 5b,c), the plume reached twice as far into the Nansen Basin (~90 km), but maximum particle loads decreased to 2.6 ppm. Successive mixing 10,12 decreased BSBW signals, whereas particle discharge towards the seafloor reduced the particle load and its association with BSBW with increasing propagation distance. East of Vilkitzky Strait (Extended Data Fig. 5d,e), ~420 km downstream the first observation, maximum particle signals decreased to 1 ppm while the association with decreasing BSBW signatures reduced further. Ongoing degradation and remineralization, evident in the observations of increased zooplankton abundances (Extended Data Fig. 1f) and heterotrophic indicator species and ammonium oxidizing Nitrococcacceae in genomic sequences of particle-associated microbes within BSBW (Supplementary Table 1) further reduced the POC load and caused discharge of dissolved carbon and nutrients into surrounding waters. Finally, north of the central Laptev Sea (Extended Data Fig.  5f,g), distinct BSBW signals were absent, and maximum particle loads of 0.4 ppm were almost exclusively located below the 27.97 kg m −3 isopycnal at ~1,500 m. BSBW-driven particle injection thus represents a pathway for organic matter from the shelf over ~1,000 km into the Eurasian Basin (Fig. 3) and provides a food source for the deep sea community, which previously could not be explained by vertical export of local sources alone 9 .

Quantifying BSBW-driven deep carbon injection
To estimate the spatial POC distribution, we scaled up our high-resolution sensor data with discrete size-fractionated POC measurements (Extended Data Fig. 6 and Extended Data Table 2). These calculations revealed maximum POC concentrations of 26.5 mg C m −3 in the plume north of Severnaya Zemlya (Extended Data Fig. 7), with a contribution of ~2/3 by the particle size fraction <0.1 mm (on average 12.4 mg C m −3 and 7.0 mg C m −3 for <0.1 mm and >0.1 mm, respectively). We further calculated the along-slope lateral POC flux based on the BSBW current velocity of 0.063 m s −1 on the day of sampling as measured by Acoustic Doppler Current Profiler and Rotor Current Meter within this transect (Methods). Calculations yielded average total lateral carbon fluxes of 84.3-117.6 g C m −2 d −1 within BSBW (<0.1 mm: 56.7-72.7 g C m −2 d −1 , >0.1 mm: 27.6-45.1 g C m −2 d −1 ), while integration over the vertical BSBW area of 19.5 km 2 resulted in a total lateral transport of 2.06 kt C d −1 . Including particles >0.1 mm, which were observed to slowly sink out of the BSBW, increased the lateral flux to 2.33 kt C d −1 . This value is equivalent to 8.55 kt CO 2 , which equals the daily emission of 2-3 average-sized coal-fired power plants e.g. as reported for China (3.6 kt CO 2 d −1 (ref. 33 )).

Temporal trends and implications for polar carbon budgets
Seasonal and inter-annual changes in carbon fixation 34,35 and BSBW production 10,13,36 could cause fluctuations of the injected POC into the deep sea. To resolve those trends, we applied the biogeochemical Regulated Ecosystem Model (REcoM2 37 ) coupled to FESOM. The model was evaluated by comparing the spatial POC distribution from the field with the respective values from the model, showing that it successfully reproduced the process of dense water formation and subduction, driving particle injection and transport towards north of Severnaya Zemlya (Extended Data Fig. 8e,g). However, the location of the modelled POC plume was ~200 m shallower, and concentrations were generally approximately ten times lower (~2.5 mg C m −3 ) than the field observations. This discrepancy could be explained by the temperature bias in the underlying FESOM model, causing a reduced density and export depth of the produced bottom water but also a reduced injection flux. Additionally, REcoM2 does not include resuspension processes, which probably play an important role for carbon transport on the shelf and might lead to an underestimation of the transported POC load. These results indicate that the process of dense water-driven particle injection could qualitatively be well reproduced by REcoM2, making it a suitable tool to resolve temporal trends of carbon outflow.
Model results revealed that pulses of carbon-rich dense water leave the shelf during and following the productive season in the second half of every year (Fig. 4). Compared with the field-based POC flux north of Severnaya Zemlya (2.33 kt d −1 ), the respective modelled flux for August 2018 was approximately ten times lower (~0.2 kt C d −1 , factor 11.68) but only approximately five times lower when using September outputs (0.52 kt C d −1 , factor 4.46). The limited temporal coverage of our field data did not allow an evaluation of seasonal outflow patterns, and in turn, a possible mismatch of POC transport peaks cannot be excluded. Moreover, in periods between POC outflow peaks, modelled fluxes almost reached undetectable levels, but resuspension of deposited material not represented in the model may cause limited fluxes to occur in between bloom phases in reality. Annual flux estimates including seasonal fluctuations were calculated by correcting modelled annual fluxes with daily flux discrepancy factors between model and field estimations. North of Severnaya Zemlya, this estimation revealed 0.26-0.67 Mt C yr −1 , while in total up to ~1 Mt C yr −1 might get transported out of St. Anna Trough (0.35-0.94 Mt C yr −1 ). This value would increase the total POC sequestration of the Barents Sea by ~1/3, considering a previously assumed burial of 2.8 Mt C yr −1 in shelf sediments only 35 . Moreover, it is equivalent to 3.6 Mt CO 2 yr −1 or ~3% of the total Barents Sea CO 2 uptake 14,38 , which would retain the CO 2 emission of a low-emitting country, such as Iceland (3.7 Mt CO 2 yr −1 ) (ref. 39 ) from the atmosphere on millennial time scales 40 due to the slow circulation of Arctic deep waters. In addition, elevated fluxes of dissolved organic carbon (DOC), which were not quantified in the field, also could be observed in the REcoM2 results (Fig. 4b), where annual total DOC fluxes fluctuated between 50% and 200% of the annual POC fluxes (Fig. 4c), promoting the role as a carbon sink.
Our FESOM model results revealed a relatively stable annual dense water outflow volume of ~2 Sv out of St. Anna Trough between 1980 and 2018 ( Fig. 4a). At the same time and location, REcoM2 results showed the POC outflow to be slightly increasing (~0.05 Mt C yr −1 to ~0.07 Mt C yr −1 ). This development could be explained by a higher biological production due to increased light and nutrient availability as a result of less sea ice cover and stratification that already more than doubled between 1989 and 2017 as reported based on satellite data 41 . With ongoing reduction of sea ice cover, however, dense water formation on the shelf also might decrease in the future 36 , which would reduce the amount and the injection depth of carbon drastically and, in turn, also the retention time. Although we could not report this scenario in our model, negative feedback effects, and hence an amplification of global warming, cannot be excluded because underlying field measurements are scarce in this region.
We could show that the co-location of dense water formation and transport with elevated biological production in the Barents Sea impacts the carbon budget of the Arctic Ocean by retaining substantial amounts of carbon in the deep Nansen Basin while providing a food source for deep sea organisms. The Barents Sea, however, is not the only shelf region with such co-location. For example, the Sea of Okhotsk, the Bering Sea or around northern Greenland 19,42 and, in particular, the Antarctic Ross 43 and Weddell seas 44-46 are characterized by dense Article https://doi.org/10.1038/s41561-022-01069-z water formation and primary production, and it is therefore probable that high carbon injection rates also occur there as well. Moreover, high-density bottom waters produced around Antarctica 47 can lead to even deeper injection with longer retention times than reported here for the Arctic Ocean. Polar regions are currently changing dramatically, and feedback effects of global warming might lead to even higher global mean temperatures in the future 48 . The scarcity of measurements in dense water outflow regions leads to uncertainties about injection systems within polar regions. A future reduction of dense water formation and hence a reduced transport of carbon from the atmosphere could result in a cumulative effect and a so far unconstrained amplification of global warming, which urgently calls for detailed studies of such systems.

Online content
Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41561-022-01069-z.

Cruise description and sensor set-up
Data and samples were collected during the research campaign ARC-TIC2018 in the Nansen Basin and Laptev Sea onboard RV Akademik Tryoshnikov from 18 August to 29 September 2018. Conductivity, temperature and depth profiles (CTD; SBE 9plus, Sea-Bird Electronics) including fluorometer (ECO Fl, Wetlabs) and turbidity data (ECO BB, Wetlabs) were acquired at each station 49,50 . An Underwater Vision Profiler (UVP 5hd; Hydroptic) was mounted on the CTD frame and acquired particle data and images with a maximum frequency of 20 Hz during descent.

Size-fractionated particulate organic carbon and nitrogen sampling and measurements
Size-fractionated POC and particulate organic nitrogen (PON) sampling was accomplished by filtering up to 60 l of sample water directly from the Niskin bottles through 0.1 mm nylon mesh filters (Millipore). Subsequently, material >0.1 mm on the mesh filters was rinsed off using sterile filtered seawater (0.2 µm) and filtered onto 0.8 µm GFF filters (Whatman). 2 l of the corresponding filtrate from 0.1 mm nylon mesh filters was filtered onto 0.8 µm GFF filters (Whatman) to assess the size fraction <0.1 mm. To reduce fragmentation of fragile aggregates >0.1 mm into the filtrate and thus the particle pool <0.1 mm, water flow during mesh filtration was adjusted to minimum. Moreover, material was rinsed off from the mesh very thoroughly to avoid loss of material >0.1 mm. Samples were stored at −80 °C before removing carbonates by acidification and tin encapsulation. Measurements of C and N content of the processed samples were carried out using an element analyser (Euro EA 3000 Elemental Analyser, Eurovector).

Size-fractionated DNA sampling for amplicon 16S and 18S rRNA gene polymerase chain reaction and sequencing
Size-fractionated DNA sampling was accomplished analogously to POC and PON sampling. The particle size fraction >0.

Analyses of amplicon sequence data and statistical analysis
Amplicon sequence variant (ASV) tables were constructed using the DADA2 pipeline v.  57 ) with the Silva database v. 132 as a reference database. 18S rRNA gene amplicons were assigned using the 'feature-classifier' v. 2019.7.0 (from package 'q2-feature-classifier' v. 2019.7.0) in QIIME 2 (ref. 58 ) and the pr2 database v. 4.12 (ref. 59 ) as a reference database.
The ASV tables were Hellinger transformed to stabilize the variance in the sequence count data for beta dispersion and PERMANOVA analyses using the decostand() function in zCompositions (rowsum cut-off = 3). Environmental metadata were z-scored (mean of data variable shifted to 0) for comparable metadata analysis.
To examine microbial community dissimilarity between groups ('outside plume', 'plume'), we performed a principal coordinate analysis (PCoA); we calculated beta dispersion of sites based on Bray-Curtis dissimilarity analysis of Hellinger-transformed ASV tables. PCoA eigenvalues were used to examine variation captured in PCoA axes. Differences of microbial communities between groups were tested with a permutational MANOVA (PERMANOVA 60 ) on the Hellinger-transformed ASV tables using the adonis2() function in vegan (v. 2.5.6).

UVP particle and image data processing
Pre-processing of particle and image data from the UVP was accomplished using the software Zooprocess 61 . Calculation of particle volume per sample volume and data binning into 5 m depth bins and standardized, evenly spaced size classes between 0.102 mm and 2.05 mm ESD (further 0.1 mm and 2 mm) on a natural logarithmic scale was carried out using standard procedures 62 . Classification and validation of acquired images (>0.8 mm) was accomplished using the web-based application Ecotaxa 63 . Images were individually classified with the assistance of machine learning classifiers into respective groups of living zooplankton and non-living marine snow. Marine snow was further classified into distinct groups using k-means clustering of principal component analysis (PCA) coordinates of morphological features of the detritus images as described elsewhere 28 .
Maps were created using Ocean Data View 64 and the International Bathymetric Chart of the Arctic Ocean (IBCAO 3 (ref. 65 )). The 3D overview of the particle volume distribution in the Nansen and Amundsen basins shown in Fig. 3 was created in PyGMT v. 0.6.1 (refs. 66,67 ) using the SRTM15+ v. 2.4 grid 68 and surface gridding 69 .

Estimating POC concentration using turbidity, fluorescence and optical particle data
To visualize high-resolution POC concentrations and to calculate total POC flux throughout the bottom water plume, we scaled up size-fractionated POC measurements using CTD-based fluorescence and turbidity data and optical particle concentration and size distribution data from the UVP. Regressions were carried out using the statsmodels (v. 0.12.2) Python module 70 .
High-resolution POC concentrations within the UVP size fraction >0.1 mm ESD were estimated using the pattern of the particle size distribution, total particle volume and POC concentration of the sampled fraction >0.1 mm (n = 92). Only particle volumes of size classes up to 2 mm ESD were included in the estimation to eliminate rare but massive outliers, for example, due to large zooplankton. The pattern of the particle size distribution was described by the two coefficients a and b of the power law function (f (x) = ax b ) fitted over the relative cumulative particle size class distribution (cumulative volume: total volume; Vol cum :Vol tot ) using ordinary least squares (OLS) regression (Extended Data Fig. 6a,b). To optimize regression and to enable zero-based power law fitting, centres of the respective ESD size bins (in µm) were used and transformed before regression (subtracted by the smallest size bin centre value of 115 µm; ESD t ) to eliminate the offset below the UVP detection limit. Measured POC concentrations >0.1 mm ESD where then correlated to the total particle volume 0.1-2 mm ESD and the calculated coefficients a and b (Extended Data Fig. 6c). The regression was achieved using a generalized linear model based on the gamma distribution model family, logarithmic link function and iteratively reweighted least squares (Extended Data Table 2). Resulting intercept and coefficients were then used to estimate POC concentrations Article https://doi.org/10.1038/s41561-022-01069-z >0.1 mm for each UVP depth bin, following OLS regression of the respective zero-transformed cumulated volumetric particle size distribution as described above.
High-resolution POC estimation of the fraction below the UVP size threshold of 0.1 mm was achieved by correlating the POC concentrations of the sampled fraction <0.1 mm ESD (n = 92) with CTD-based turbidity and chlorophyll a fluorescence sensor data of the respective sampling location and depth (Extended Data Fig. 6d). Therefore, a robust linear regression model based on iteratively reweighted least squares including Andrewʼs Wave M-estimator as weight function was used (Extended Data Table 2). The resulting intercept and coefficients of this regression were then used to estimate POC concentrations <0.1 mm ESD for each CTD data point.
Fitting of the resulting estimates of the POC load <0.1 mm and >0.1 mm ESD are presented as the difference between the measured and estimated POC concentration for the respective sampling location and depth (Extended Data Fig. 6e). Almost all larger outliers were located in the euphotic zone, pointing towards non-matching locations of sampling and data acquisition due to the patchiness of the POC distribution in this zone. Nevertheless, about 80% of the estimates showed a difference below 10 mg m −3 , representing a relatively robust regression considering the different nature of the approaches, the fact that data acquisition of CTD and UVP occurred during descent and POC sampling during ascent with wind drift in between and that no organisms were picked out of the samples.

Definition of BSBW, current velocities and data processing
BSBW was defined based on the transformation characteristics of Atlantic waters of the West Spitsbergen current region. Being denser than waters in the intermediate density layer and less saline than dense Atlantic water from Fram Strait, transformed BSBW north of Severnaya Zemlya was characterized by absolute salinities <35.06 g kg −1 and potential density anomalies >27.97 kg m −3 after leaving the Barents Sea shelf through St. Anna Trough. Units used for definition were based on the Thermodynamic Equations Of Seawater-2010 71 and calculated using, for example, the Python package gsw (v. 3.4.0).
The average ocean current velocity of the area covered by BSBW across the westernmost transect was estimated from Acoustic Doppler Current Profilers (ADCPs, Teledyne RDI) and rotor current meters (RCM) deployed in a high-resolution mooring array. The array consisted of seven moorings (AK1 to AK7) arranged perpendicular to the continental slope in water depths between 300 m (AK1) and 3,015 m (AK7) from August 2015 to September 2018 (E.R.-C. et al., manuscript in peer review). The sampling frequency for ADCPs and RCMs was 90 minutes and 120 minutes, respectively. The effects of the tide and inertial periods were removed with a six-day running average window, and temporal resolution was reduced to one day. Only velocity measurements located within the BSBW region were used (AK2 to AK5). RCM current records at AK4 and AK5 stopped on 31 January 2016 and 11 May 2017, respectively. These gaps in the time series were recovered by scaling velocities from the ADCPs. Before the current records stopped, correlation between time series from both instruments for the zonal and meridional components were 0.85-0.71 and 0.56-0.47 for AK4 and AK5, respectively. Currents from 20 August 2018 were aligned in the along-slope direction and used to represent the time period when the hydrographic transect was occupied. The velocity records in the area where BSBW occurred were horizontally interpolated and averaged. More detailed descriptions of deployment, water mass definition, calculation procedures and the resulting dataset will be published separately (E.R.-C. et al., manuscript in peer review).

Estimating BSBW-driven lateral POC flux in the westernmost transect closest to the St. Anna Trough outflow
BSBW-driven lateral POC flux was estimated for the westernmost transect closest to St. Anna Trough based on POC load estimations described above starting with the station-wise integration of POC contents within BSBW per m 2 water column. For each station of this transect, POC contents per m 2 BSBW per station were then multiplied by half of the distance to the nearest stations north-and southwards, respectively. No extrapolation towards the south of the southernmost and towards north of the northernmost station was calculated to perform a conservative flux estimation. The resulting POC content in the whole vertical area of BSBW was then multiplied by the BSBW current velocity (0.063 m s −1 ) within the same transect and from the day of sampling (above). Finally, calculation of different time intervals (day, year) and average lateral flux per m 2 followed.
On the basis of the assumption that a fraction of POC > 0.1 mm ESD was transported into the area by BSBW but was dislocated due to gravitational settling (Fig. 1, Extended Data Fig. 1 and Extended Data Fig. 7), one can conclude that this material was also transported with the BSBW current velocity of 0.063 m s −1 . To obtain estimates of the potential additional fluxes by this material, we calculated the possible additional POC load as follows: maximum depth with increased POC concentration >0.1 mm below BSBW was calculated by identifying the depth for each station where ambient concentrations below BSBW were reached. Ambient concentrations were defined for each station as the average POC load >0.1 mm within 100 m above the BSBW subtracted by its standard deviation. Ambient concentrations were reached right below BSBW in all stations except for one station located at 82.11° N and 94.84° E. For this profile, POC concentrations >0.1 mm ESD were integrated over the layer below the depth of the last BSBW signal and the identified maximum depth. This POC load per m² water column was then multiplied by half the distances to the next two stations north-and southwards as described above and multiplied by the BSBW current velocity of 0.063 m s −1 .

FESOM modelling of current speeds and particle backtracking
To better understand the velocity structure in the northern Barents Sea, we used the velocity field from the Finite-Element Sea ice-Ocean Model v. 1.4 (FESOM 1.4 (ref. 18 )). FESOM 1.4 is based on triangular unstructured meshes for both the ocean and sea ice components. The global model grid has 4.5 km resolution in the Arctic Ocean and 24 km in the North Atlantic. The model configuration used here was forced with atmospheric reanalysis data from JRA55-do v. 1.3 (ref. 72 ) and reasonably represents the hydrography and velocity structure in the Arctic Ocean 73,74 . One should note that in the model, a known warm bias of around 1 °C relative to the Polar Science Center Hydrographic Climatology (PHC v. 3 (ref. 75 )) was present in the Eurasian Basin of the Arctic Ocean ( Fig. 4 in ref. 73 ). Resulting water mass distribution, however, was similar to field observations, although the location of the dense water plume at the westernmost transect at ~90° E was slightly shallower ( Fig. 1 and Extended Data Fig. 2). Following release depths for particle backtracking were adjusted accordingly to match the observed co-location of particle maxima with dense water.
A Lagrangian particle-tracking algorithm was used to determine the origin of representative particle classes observed at the continental shelf break. It has been successfully applied before to study the catchment area of sediment traps deployed in Fram Strait 76 Fig. 2c). Release was carried out every 14 days in 2018 and back tracked until they reached the ocean surface, resulting in 1,680 trajectories. A time step of 30 min was used for the trajectory calculations, yielding bi-hourly positions. The horizontal displacement of particles was computed with daily-averaged velocity fields from FESOM 1.4, and the vertical displacement was determined from constant sinking velocities measured for particle samples from Fram Strait 78 . The sinking velocity (v) https://doi.org/10.1038/s41561-022-01069-z was computed with its empirical relationship to particle ESD described by the coefficients a and b of the power law function (v = a(ESD) b ). We conducted backtracking model runs with sinking velocities representing non-ballasted (a = 34.6, b = 1.39), small (0.08 mm ESD) and large (0.5 mm ESD) particles (v = 1.03 m d −1 and 13.2 m d −1 , respectively) and small ballasted particles (a = 94.58, b = 0.96) with 0.08 mm ESD (v = 8.58 m d −1 ). Selected numerical results are shown in Extended Data Table 1.

Evaluation of the injection process and seasonal and inter-annual trends using REcoM2
The Regulated Ecosystem Model (REcoM2) is a biogeochemical model describing the lower trophic levels of the marine Arctic ecosystem with one zooplankton class, the phytoplankton types diatoms and nanophytoplankton and one type of detritus 37 . It includes descriptions of the cycles of nitrogen, silicon and iron and the carbon cycle, whereas water mass and current properties are based on FESOM (above). The FESOM-REcoM2 set-up was run from 1980 to 2018. The first 25 years of the run have been validated by Schourup-Kristensen et al. 79 , while the time-varying trends in Arctic nitrate supply are discussed in Oziel et al. 80 .

Combining model and field data to estimate annual POC flux
To provide realistic annual POC flux estimations including seasonal outflow patterns, outputs from the REcoM2 model runs were corrected using flux calculations from the field (above). A correction factor for the model was created by dividing the monthly RECOM2 output of the average daily POC flux within BSBW of the westernmost field transect north of Severnaya Zemlya by the respective daily field data-based flux. Due to the fact that field data collection took place at the end of August (22 August) and that we could not exclude a temporal mismatch between peaks of field and model-based POC outflow from the shelf, we calculated this correction factor for August and September 2018, respectively. Annual REcoM2-based POC fluxes for the transects north of Severnaya Zemlya and St. Anna Trough were then multiplied by those two correction factors, resulting in a range of field data-corrected POC fluxes for those transects.

Data availability
CTD data are available on PANGEA 49 and the Arctic Data Center 50 . Detailed UVP image and particle data that support the findings are publicly available on EcoTaxa (https://ecotaxa.obs-vlfr.fr, project number 2463) and Ecopart (https://ecopart.obs-vlfr.fr; project number 239). Results of the PCA-based k-means clustering of detritus classes, depth-binned distributions of zooplankton and particle size classes and the measured POC/PON data and the resulting POC estimates are publicly available on the PANGAEA data repository (

Code availability
Results were achieved using publicly available Python and R packages as mentioned in the Methods section. All scripts for POC estimations can be provided upon request. Article https://doi.org/10.1038/s41561-022-01069-z Extended Data Fig. 3 | Results of the principal coordinate analysis of the particle associated community composition between plume and non-plume (PERMANOVA, Hellinger transformed, 999 permutations). a, prokaryotic community differences (based on 16S rRNA genes) associated with particles <0.1 mm ESD. b, prokaryotic community differences (based on 16S rRNA genes) associated with particles >0.1 mm ESD. c, eukaryotic community differences (based on 18S rRNA genes) associated with particles <0.1 mm ESD. d, eukaryotic community differences (based on 18S rRNA genes) associated with particles >0.1 mm ESD. Weak relationships could be identified between plume and nonplume particle-associated microbes within sinking particles (P = 0.019, R 2 = 0.102 for prokaryotes, and P = 0.032, R 2 = 0.082 for eukaryotes), but correlations were stronger within suspended particles (P = 0.007, R 2 = 0.123 for prokaryotes, and P = 0.002, R 2 = 0.117 for eukaryotes).