A total of 5 years (from 2015 to 2019) of aerosol sampling and measurements were carried out at the Gruvebadet (GVB) Observatory (78.918∘ N, 11.895∘ E; 61 m above mean sea level), located at about 800 m southwest of the village of Ny-Ålesund in the Svalbard archipelago. In the northeastern direction towards the Ny-Ålesund research village, a clean area was established and motorized activity and other potentially contaminant activities were forbidden. The geographic location of GVB and the dominant winds ensure minimal anthropogenic contamination from local emissions while also capturing long-range transported pollution air masses.

The aerosol size distribution was measured by an aerodynamic particle sizer (APS; TSI model 3321, 52 channels) . The APS spectrometer provides real-time aerodynamic measurements of particles from 0.5 to 20 µm diameter. Typical single-channel uncertainty in aerosol number concentration measured by the APS is ±10 % according to the specification sheets. The aerosol size distribution was usually measured from March to September (see Fig. ), with approximately 21200 hourly observations in total from 2015 to 2019, corresponding to approximately 909 and 140 aerosol size distributions after averaging to daily and weekly observations, respectively. Aerosol light absorption was simultaneously measured at three wavelengths (i.e., 467, 530 and 660 nm) by a particle soot absorption photometer (PSAP, Radiance Research) at a 1 min time resolution and then averaged into hourly observations. The raw data from PSAP were corrected according to . Aerosol absorption coefficient is directly proportional to the concentration of equivalent black carbon (eBC) , a marker of anthropogenic and wildfire emissions in the Arctic . Details regarding the PSAP and eBC measurements at the site are described by . The APS and PSAP are attached to a same inlet, which has followed EUSAAR-ACTRIS protocol since 2011 and runs about 4 m above the ground. Further information can be found elsewhere .

Aerosol filter samples were collected daily from February 2015 to August 2018, using a Tecora SkyPost sequential sampler equipped with a PM10 sampling head, operating following the EN 12341 European protocol. PM10 was collected on Teflon filters (Pall Corp., 47 mm diameter, 2 µm porosity, collection efficiency 99.6 %); 606 daily samples were analyzed, of which 570 samples overlapped with the aerosol size distribution measurements. Samples were handled in a clean room (class 10 000), under a laminar flow hood (class 100) to minimize contamination. Half of each filter sample was extracted in an ultrasonic bath (15 min) using ultrapure water (MilliQ, 18 MΩ cm, 12 mL) for subsequent analysis of inorganic anions (Cl-, Br-, NO3- and SO42-) and cations (Na+, NH4+, K+, Mg2+ and Ca2+) and selected organic anions (methanesulfonate (MSA) and oxalate) on a three Dionex ion chromatography (IC) system equipped with electrochemical-suppressed conductivity detectors. The configuration of the three IC system is described by and . Analytic uncertainty is typically below 5 %.

Meteorological parameters, including wind speed, wind direction, relative humidity and ambient temperature, were recorded hourly by a Vaisala thermo-hygrometer model HMP45AC and a Young Marine wind sensor model Wind Monitor 05106 at a height of 10 m (a.g.l.) on the Amundsen-Nobile Climate Change Tower in the neighborhood of Gruvebadet observatory.

Cluster analysis is commonly used to interpret particle number size distribution measurements . As we mainly focus on supermicrometer aerosol, we conduct cluster analysis of the aerosol volume size distribution rather than aerosol number size distribution using a k-means clustering algorithm following a standard procedure by . The underlying principle of the k-means clustering is to minimize the sum of squared Euclidean distances between each data point and the corresponding cluster centers (the average spectra of the subgroup). This allows for dividing the whole dataset into a predefined number of subgroups that are as different as possible from each other and as coincident as possible within themselves.

To cluster the spectra modes irrespective of the magnitude of the concentrations, the aerosol size distributions were pre-processed using min–max normalization: 1yi=xi-min⁡1≤j≤n{xj}max⁡1≤j≤n{xj}-min⁡1≤j≤n{xj}, where x and y are the absolute and normalized concentrations, respectively. The subscript i and j denote size bins, and n is the number of size bins. The normalized concentrations were then clustered using k-means.

Source apportionment of sulfate was carried out following to understand potential sources of sulfate for the identified aerosol clusters. We first apportion Na+ and Ca2+ to sea salt and non-sea salt fractions. In general, Na+ and Ca2+ are considered proxies for sea salt and mineral dust, respectively. However, a small amount of Na+-containing aerosol in remote regions could be brought in by non-sea salt sources and some fractions of Ca2+-containing aerosol originate from sea spray . Assuming a Ca2+ / Na+ ratio (w/w) of 0.038 for sea water and 1.78 for mineral dust , the measured total Na+ and Ca2+ can be apportioned to a sea salt (ss) fraction and a non-sea salt (shortened to nss) fraction calculated by

2ss-Na+=Na+-nss-Ca2+/1.78,3nss-Ca2+=Ca2+-0.038×ss-Na+.

The sea salt fraction of sulfate (ss-SO42-) was calculated by multiplying the ss-Na+ (as a sea spray marker) concentration by 0.253 (indicating the SO42- / Na+ w/w ratio in sea water) . The non-sea salt fraction of sulfate (nss-SO42-) was calculated by subtracting the ss-SO42- contribution from the total SO42- concentrations. The mineral fraction of sulfate (mineral-SO42-) was estimated by multiplying the nss-Ca2+ content by 0.59 (SO42- / Ca2+ w/w ratio in the uppermost part of Earth's crust) . It is worth noting that Ca in high-latitude dust may not be in the form of carbonate and thus will not be measurable as Ca2+ . Some studies also reported a low amount of Ca-containing high-latitude dust . Thus, the nss-Ca2+ is a conservative tracer for mineral dust. Sulfate formed by atmospheric oxidation of dimethylsulfide (DMS) from micro-algae, bio-SO42-, was estimated by multiplying methanesulfonate (MSA) concentration by 3.0, which was derived from the relationship between MSA and SO42- at the same site by . In their study, they used 136 PM10 samples from the spring–summer season of 2014 to establish the MSA and bio-SO42- relationship. The anthropogenic fraction of sulfate (anthr-SO42-) in each sample is then estimated by subtracting the sum of the ss-SO42-, mineral-SO42-, and bio-SO42- fractions from the total SO42- concentration.

The Hybrid Single Particle Lagrangian Integrated Trajectory (HYSPLIT 5.0.0) model was used to calculate hourly backward trajectories arriving at an altitude of 100 m (average mean sea level) at the GVB from 2015 to 2019. The determination of the length of the back trajectories is a compromise between the uncertainty in calculation, which increases with time, and the typical lifetime of aerosol in the Arctic, which is usually between 3 and 8 d for black carbon and up to 39 d during Arctic haze . In the present study, a 7 d air mass backward trajectory was chosen.

The HYSPLIT back trajectories provide a detailed history of the air parcels, including latitude, longitude, altitude and the height of the mixing layer (ML). The back trajectories are divided into above or below the mixing layer height. Daily surface types in the Northern Hemisphere were obtained from the National Ice Center's Interactive Multisensor Snow and Ice Mapping System (IMS) , National Snow & Ice Center. The IMS map of the Northern Hemisphere classified each of 1024 × 1024 grid cells (24 km spatial resolution) as sea, land (without snow), sea ice or snow-covered land (snow). The daily surface types are temporally allocated with the back trajectories. Each data point from the hourly back trajectories is labeled as (1) sea, (2) sea ice, (3) snow, (4) land or (5) above ML.

The potential source contribution function (PSCF) values were calculated using the “openair” package in R to identify the potential geographic origins of each aerosol cluster. The geographic region covered by the back trajectories was divided into 1∘ × 1∘ grid. The PSCF value is calculated by 4PSCFi,j=mi,j/ni,jWi,j, where ni,j is the total number of endpoints that fall in the i,jth cell, and mi,j is defined as the number of endpoints in the same cell that exceed the threshold criterion. In the present study, the 75th percentile of total volume concentration for each cluster was used as the criterion value. A higher ratio of mi,j/ni,j indicates a higher probability of a particular grid through which a passing air parcel would result in a higher concentration at the receptor . Wi,j is an empirical weight function to reduce the uncertainty of mi,j/ni,j for the cells with ni,j less than 2 times the grid average number of the trajectory endpoint (2×nave). 5Wi,j=1.0,ni,j>2nave0.75,nave<ni,j≤2nave0.5,0.5nave<ni,j≤nave0.15,0<ni,j≤0.5nave

The k-means cluster analysis of aerosol volume size distributions was performed using 909 daily spectra collected from 2015 to 2019. The determination of the most appropriate number of clusters is critical. Here, the Dunn Index (DI) and silhouette width (SW) were calculated for cluster numbers ranging from 2 to 15 (see Fig. ). DI is to identify sets of clusters that are compact, with a small variance between members of the cluster, and well separated, where the cluster centers are sufficiently far apart, as compared to the within cluster variance. A higher DI indicates better separated clusters. SW refers to a method of interpretation and validation of consistency within clusters of data. A high SW indicates that an individual spectrum is well matched to its own cluster and poorly matched to neighboring clusters. For a given assignment of clusters, higher DI and SW values indicate better clustering results. According to the DI and SW values versus cluster numbers (Fig. ), five- and seven-cluster solutions appear to be the two promising solutions. For the seven-cluster solution, we observed an obvious increase in the DI value but a decrease in the SW value compared to those from the five-cluster solution, suggesting that some of the clusters in the seven-cluster solution tend to be similar to neighboring clusters. Figure  illustrates that the C3 to C6 aerosol categories overlapped with each other and are not well separated, especially when taking into account the standard deviation. The five-cluster solution is able to overcome the overlapping issues and thus was chosen in the present study.

Note that the five clusters (C1, C2, C3, C4 and C5) were ranked by the main aerodynamic diameter of the peak concentration in the average spectrum from small to large diameters for C1 to C5 (Fig. ). Figure  shows the daily temporal trends and the polar plots of the five aerosol clusters. The aerosol types are characterized as follows.

C1, which occurs 9.3 ± 5.1 % (μ ± σ, average from monthly occurrence frequency) of the time. C1 presents a bimodal size distribution with a dominant peak in the accumulation mode at ∼ 0.7 µm and a lower peak at ∼ 3 µm (Figs. a–b and a). The predominant months were March (15.9 %) and April (18.8 %), i.e., in spring (Fig. c). The relative contribution from different regions to back-trajectory air parcels for C1 is similar to that in March and April, which is characterized by the highest fraction of historical air parcels above mixing-layer height among the five clusters or all seasons (Fig. a).

The occurrence of the five clusters and the key chemical species associated with each cluster are summarized in Table . Note that the relative occurrences of the clusters are limited to the study period (mainly from March to October), and thus with limited coverage of the winter season. The interannual variations of fraction of occurrence from March to June are also shown in Fig. d–g. No clear interannual patterns were observed for the aerosol clusters identified. The polar plots show that high concentrations of the five clusters were typically found when wind speed was higher than 5 m s-1 (Fig. a), which is consistent with the probability density distribution of the wind speed for aerosol volume concentrations exceeded their 90th-percentile values within each cluster (Fig. ).

Figures  and present average concentrations of each chemical species and absorption coefficients for the identified clusters, respectively. Absorption coefficients (Fig. b), oxalate (Fig. j), total SO42- (Fig. h) and NH4+ (Fig. f) were enhanced for C1. These components are also enriched in C2, along with Na+ (Fig. d), Cl- (Fig. e) and Br- (Fig. i). Absorption coefficients were relatively low for C3–5 compared to C1 and C2 (see Fig. b). MSA (Fig. k), Na+ (Fig. d) and Cl- (Fig. e) were enriched in C3. C4 showed an enhanced concentration in MSA (Fig. k), Na+ (Fig. d) and Cl- (Fig. e), whereas C5 was characterized by lowest K+ (Fig. a), NH44+ (Fig. f), NO3- (Fig. g) and SO42- (Fig. h). The relationship between C1–5 and the absorption coefficient is shown in Fig. b. Clearly, the absorption coefficient for C1–2 was enhanced in comparison to that for C3–5. We will discuss these results in detail in Sect. .

Na+ and Ca2+ are usually apportioned to a sea salt and a non-sea salt fractions using characteristic ratios of Ca2+ and Na+ in sea water and in the Earth's crust (Eqs. –). Figure  shows a scatter plot of Ca2+ versus Na+ for the clusters, including Ca2+ / Na+ ratios for sea water (0.038, w/w) and the Earth's crust (1.78, w/w). The regression line of Na+ and Ca2+ for C5 is close to the “mineral” ratio, whereas it is closer to the sea water ratio for C2–4. Note that the Na+ and Ca2+ regression line for C3 fits very well with that in sea water. The regression line for C1 is in the middle of the mineral ratio and the sea water ratio.

The apportioned ss-SO42-, mineral-SO42-, bio-SO42-, anthr-SO42- and sea salt aerosol (Fig. ) further improve our understanding of the aerosol cluster-associated sources. C2–4 were enhanced with respect to ss-SO42-, ranging from 66.1 ng m-3 in C2 to 78.9 ng m-3 in C4, as well as SSA (ranging from 853.4 ng m-3 in C2 to 1019.2 ng m-3 in C4). A high bio-SO42- fraction was observed in C3 (86.2 ng m-3) and C4 (61.6 ng m-3), whereas mineral-SO42- is elevated in C1 (7.7 ng m-3) and C5 (7.9 ng m-3). The anthr-SO42- level decreased from C1 (763.8 ng m-3) to C5 (49.7 ng m-3). Anthr-SO42- contributes to 34.4 % in C5 to 91.7 % in C1.

The PSCF probability map (Fig. ) for C1 illustrates that air mass origin was mainly from northeastern Russia (lat > 60∘). Potential geographical source regions of C2 are more complex and likely distributed in northern Eurasia (lat > 60∘), the western and northern coasts of Greenland, the Norwegian Sea, the Barents Sea and the North Sea (lat > 55∘). In contrast to C1–2, several regions of elevated probability for C3 appear in open-sea and open-ocean regions (lat > 55∘), including the Norwegian Sea, the North Atlantic Ocean, the Barents Sea, the Kara Sea and Baffin Bay. C4 is likely associated with both land and snow (e.g., Greenland, northern Europe, northwestern Russia) and sea and ocean regions (e.g., the Greenland Sea, the Norwegian Sea, the North Atlantic Ocean, the Barents Sea, Baffin Bay). C5 exhibits high PSCF values near the western and northern coasts of Greenland (e.g., the Kangerlussuaq area), northern Eurasia, and northern Alaska. For the whole aerosol population, high PSCF values were broadly observed in the surrounding sea and ocean regions and occasionally observed in land and snow regions (e.g., northern Eurasia and Greenland). Note that potentially local sources can not be captured by the back-trajectory analysis.

The dominant source for each aerosol cluster that distinguishes the aerosol clusters from each other was identified based on (1) volume size distribution, (2) chemical components, (3) seasonality in occurrence of the clusters and (4) potential source maps. It is possible that each cluster does not represent a pure source but contain signals of other aerosol sources​​​​​​​. In the following contents, sources of aerosol clusters refer to the sources that distinguish the aerosol clusters from each other.

In the present study, Arctic aerosol is characterized by prevalent Arctic haze between winter and late spring, biogenic aerosol and mineral dust during warm seasons and sea spray aerosol throughout all seasons. The seasonal cycle in the Arctic aerosol is driven by many aspects, such as the annual cycle in Arctic sea ice, temperature, radiation, atmospheric oxidants, cloud properties, and seasonally varying transport and removal mechanisms . In this study, potential drivers of the coarse-mode anthropogenic aerosol and natural aerosol are discussed in Sect.  and .