Concentrations of INPs as a function of size were measured with the micro-orifice uniform deposit impactor-droplet freezing technique (MOUDI-DFT; Mason et al., 2015b). This technique involves collecting size-fractionated aerosol particles on hydrophobic glass slides with a micro-orifice uniform deposit impactor (MOUDI; Marple et al., 1991) and determining the freezing properties of collected aerosol particles with the droplet freezing technique (DFT). Details are given below.

The combination of an aerodynamic particle sizer (APS) and a scanning mobility particle sizer (SMPS) was used to measure the aerosol number and surface area as a function of size. The APS (model 3321, TSI, Shoreview, MN, USA) measures diameters using the time-of-flight technique (Baron, 1986). At all three sites, the APS was operated with a sample flow of 1 L min-1 and a sheath flow of 4 L min-1. The aerodynamic diameter range measured by the APS was 0.54–20 µm. Due to possible drop-off in the sampling efficiency of the APS at sizes below 0.7 µm (Beddows et al., 2010), only APS data at sizes above 0.7 µm are used here, as done previously (Maguhn et al., 2003). The SMPS measures diameters based on the mobility of a particle in an electric field (Asbach et al., 2009; Hoppel, 1978). The SMPS was equipped with an inertial impactor at the inlet that removed large particles outside the measurement range. At Amphitrite Point, the SMPS (model 3936, TSI) was operated at a 0.57 L min-1 sample flow with a 2 L min-1 sheath flow and was used to measure particles with mobility diameters from 18.4 to 930.6 nm. At Labrador Sea and Lancaster Sound, the SMPS (model 3034, TSI) was operated at a 1 L min-1 sample flow rate with a 4 L min-1 sheath flow and was used to measure particles with mobility diameters from 10 to 487 nm. The sampling condition and strategy are discussed below for each site.

Sampling occurred at three coastal marine sites: Amphitrite Point (48.92∘ N, 125.54∘ W) on Vancouver Island in British Columbia, Canada; Labrador Sea (54.59∘ N, 55.61∘ W) off the coast of Newfoundland and Labrador, Canada; and Lancaster Sound (74.26∘ N, 91.46∘ W) between Devon Island and Somerset Island in Nunavut, Canada (Fig. 1 and Table 1). All measurements were conducted as part of the Network on Climate and Aerosols: Addressing Key Uncertainties in Remote Canadian Environments (NETCARE). Data from NETCARE are available on the Government of Canada Open Government Portal (Si et al., 2018). The sampling dates, ambient RH values, ambient temperatures, and wind speeds during sampling are summarized in Table 1. Additional details about the three coastal marine sites are given below.

At Labrador Sea and Lancaster Sound, a dryer was not used prior to sampling with the MOUDI, APS, and SMPS. Hence, for these two sites, all data correspond to the RH and temperatures during the sampling. The sizes measured by the MOUDI and the APS were aerodynamic diameters, while the SMPS measured mobility diameter. To allow for comparison between the INP data, APS data, and SMPS data at these two sites, all the SMPS data have been converted to aerodynamic diameters (see Sect. S1 for details).

At Amphitrite Point, a dryer was also not used when sampling with the MOUDI and APS. On the other hand, dryers were used prior to sampling with the SMPS. To allow for comparison between the INP data, APS data, and SMPS data at this site, a free parameter was used to convert the SMPS data under dry conditions to aerodynamic diameters at the RH and temperature during the sampling. The free parameter was determined from the optimal overlap between the SMPS and APS data. This type of approach has been used successfully in the past to merge SMPS and APS data (Beddows et al., 2010; Khlystov et al., 2004; see Sect. S2 for details).

For each MOUDI sample collected for INP analysis, a 3-day back trajectory was calculated using the HYSPLIT4 (Hybrid Single-Particle Lagrangian Integrated Trajectory) model of the NOAA Air Resources Laboratory (Stein et al., 2015). The GDAS (Global Data Assimilation System) 1∘ meteorological data were used as input. Back trajectories were initiated at the beginning of each MOUDI sampling period and at every hour until the end of the sampling period. The initiating height was the same as the height of the MOUDI sampling inlets as mentioned in Sect. 2.3. Back trajectories were also initiated at heights of 50 m and 150 m a.g.l. for each location to determine if the trajectories were sensitive to the height of initiation.

A global model of INP concentrations relevant for mixed-phase clouds was used to predict concentrations of INPs at the three sampling sites (Vergara-Temprado et al., 2017). The model considers ice nucleation by K-feldspar, associated with desert dust, and marine organics, associated with sea spray aerosols, as INPs. In this model (GLOMAP-mode), aerosol number and mass concentration of several aerosol species are simulated in seven lognormal modes (three insoluble and four soluble). The model has a horizontal resolution of 2.8×2.8∘ with 31 vertical levels, and it is run for the year 2001 with meteorological fields from the European Centre for Medium-Range Weather Forecasts (ECMWF). Model output for the year 2001 was used, since this model output was available from previous studies. The model includes a parameterization of boundary-layer turbulence (Holtslag and Boville, 1993). The aerosol components are emitted, internally mixed with the species of their mode, and several aerosol microphysical processes, including new particle formation, particle growth, dry deposition, and wet scavenging are represented (Mann et al., 2014). The INP concentrations are determined using a laboratory-based temperature-dependent density of active sites (active sites per unit surface area) for K-feldspar (Atkinson et al., 2013) and a parameterization for marine organics based on the INP content of microlayer samples (expressed as active sites per unit mass of organic carbon; Wilson et al., 2015) following the method shown in Vergara-Temprado et al. (2017) in Appendix 2.

To predict INP concentrations at the three coastal marine sites, we used the output of the model for the grid cells that overlapped with the measurement locations. Since the measurements were carried out at the surface, output from the lowest level of the model was used. We calculated the mean concentrations of INPs from K-feldspar and marine organics for the months when measurements were made. For the simulations at Amphitrite Point, Labrador Sea, and Lancaster Sound, the months of August, July, and July were used, respectively.

As mentioned above and as done previously, the model output from the year 2001 was compared with measurements from different years. The inter-annual variability of aerosol concentrations simulated in the model is expected to be up to a factor of 2 due to differences in meteorological conditions (Marmer and Langmann, 2007). Model output for the year 2001 has been found to be able to reproduce the mass concentrations of mineral dust and marine organic aerosols within an order of magnitude with observations made in various years (Vergara-Temprado et al., 2017). Furthermore, the model output for the year 2001 was able to reproduce 62 % of the INP concentrations measured from studies spanning from 1973 to 2016 within an order of magnitude, which is the uncertainty in the predicted INP concentrations reported here (Fig. 8).

Figure 2 shows the 3-day back trajectories initiated for every hour during the MOUDI sampling at the three sites. Back trajectories initiated at heights of 50 m and 150 m a.g.l. (see Figs. S1–S2) showed similar results. When considering all the back trajectories, at Amphitrite Point, 94 % of the time was spent over the ocean, at Labrador Sea, 40 % of the time was spent over the ocean, and at Lancaster Sound, 63 % of the time was spent over the ocean. The rest of the time was spent over the land. At Amphitrite Point, although the air masses were predominantly from the ocean based on the back trajectory analysis, the air masses did pass over local vegetation including coastal western hemlock. This local vegetation could potentially release enough INPs to overwhelm a small INP source from the ocean. Therefore, it was not possible to determine if the INPs are of marine or terrestrial origins based on the back trajectories alone. INPs may even have been long-range transported from sources that were not reached by the 3-day back trajectories (Vergara-Temprado et al., 2017).

In Fig. 3, the average INP number concentration is plotted as a function of size for the freezing temperatures of -15, -20, and -25 ∘C. These three temperatures were chosen because freezing events were rare at temperatures warmer than -15 ∘C, and for some MOUDI stages, all the droplets were frozen at temperatures lower than -25 ∘C, making calculations of INP concentrations using Eqs. (1)–(2) not possible at temperatures lower than -25 ∘C. Mason et al. (2015a) previously reported the average INP number concentrations as a function of size at Amphitrite Point for the time period of 6–27 August 2013. Here we report the average INP number concentrations as a function of size at the same site for a subset of the measurements (23 out of 34 samples) from Mason et al. (2015a) when both APS and SMPS data were available. Not surprisingly, the results shown here are very similar to the results shown by Mason et al. (2015a). The result for Labrador Sea shown in Fig. 3 has also been reported previously in Mason et al. (2016), while the result for Lancaster Sound is new and represents the first report of INP concentrations as a function of size in the Arctic marine boundary layer. Lancaster Sound had the lowest INP concentrations among the three sites, with average concentrations of INPs of 0, 0.16, and 0.67 L-1 for the freezing temperatures of -15, -20, and -25 ∘C, respectively. These numbers are consistent with several previous measurements reported in the Arctic. For example, Mason et al. (2016) reported the following mean concentrations at a surface site in Alert, Nunavut: 0.05, 0.22, and 0.99 L-1 for freezing temperatures of -15, -20, and -25 ∘C, respectively. Bigg (1996) reported mean INP concentration of 0.01 L-1 at -15 ∘C on an icebreaker in the Arctic. Fountain and Ohtake (1985) measured mean INP concentrations of 0.17 L-1 at -20 ∘C at a surface site in Barrow, Alaska.

At Amphitrite Point and Labrador Sea, the majority of INPs measured were >1 µm in diameter at all the temperatures studied. At Lancaster Sound, the majority of INPs were also >1 µm at -25 ∘C. At -15 ∘C, the concentrations of INPs were not above the detection limit at any of the sizes, while at -20 ∘C, freezing was only observed for sizes between 0.56 and 1 µm.

As mentioned above, the average concentrations of aerosol number and surface area as a function of size during sampling periods were determined from measurements with an SMPS and an APS. The results are shown in Fig. 4. The size distributions were consistent with the size distributions measured at a mid-latitude North Atlantic marine boundary layer site by O'Dowd et al. (2001); see Fig. S3. The average total number concentrations were 1487±512 cm-3, 3020±128 cm-3, and 946±254 cm-3 for Amphitrite Point, Labrador Sea, and Lancaster Sound, respectively. The number concentration at the Arctic site Lancaster Sound may have been influenced by new particle formation in the summer Arctic marine boundary layer (Burkart et al., 2017; Tunved et al., 2013). For the size range of measured INPs (0.18–10 µm), on average, <3 % of the number concentration was supermicron in diameter, and <47 % of the surface-area concentration was supermicron in diameter.

The ice-nucleating ability on a per number basis is represented as the fraction of aerosol particles acting as an INP. Shown in Fig. 5 is the fraction of aerosol particles acting as an INP as a function of size. To generate Fig. 5, the aerosol number concentration (Fig. 4a) was first binned using the same bin widths as the MOUDI, resulting in the total aerosol number concentration in each size bin (Fig. S4a). Then the INP concentration (Fig. 3) was divided by the aerosol number concentration (Fig. S4a). Figure 5 shows that the fraction of particles acting as an INP is strongly dependent on the size. For Amphitrite Point and Labrador Sea, and for diameters of around 0.2 µm, approximately 1 in 106 particles acted as an INP at -25 ∘C. On the other hand, at the same sites and for diameters of around 8 µm, approximately 1 in 10 particles acted as an INP at -25 ∘C. A similar trend may be present at Lancaster Sound, but at the smaller sizes investigated, the concentrations of INPs were below the detection limit. The results in Fig. 5 show that the large particles at the three sites studied are extremely efficient at nucleating ice, and as a result, even though the number concentration of large particles might be small in the atmosphere, they can make an important contribution to the total INP number concentrations.

The strong dependence on the size shown in Fig. 5 is consistent with the small number of previous studies that investigated the fraction of aerosol particles acting as an INP as a function of size. Berezinski et al. (1988) studied INPs collected at 100–500 m a.g.l. in the southern part of the European territory of the former USSR. At a freezing temperature of -20 ∘C and for a diameter of 0.1 µm, approximately 1 in 105 particles acted as an INP, while for a diameter of 10 µm, approximately 1 in 100 particles acted as an INP. A study of residuals of mixed-phase clouds by Mertes et al. (2007) found that 1 in 10 supermicron particles acted as an INP, while only 1 in 103 submicron particles acted as an INP. Huffman et al. (2013) studied INPs collected at a semi-arid pine forest in Colorado, USA. At a freezing temperature of -15 ∘C and for a diameter of 2 µm, approximately 1 in 103 particles acted as an INP, while at the same freezing temperature but for a diameter of 10 µm, more than 1 in 100 particles acted as an INP.

The surface active site density, ns, represents the number of ice nucleation sites per surface area (Connolly et al., 2009; Hoose and Möhler, 2012; Vali et al., 2015). This parameterization assumes that freezing is independent of time and can be scaled with surface area. Although these assumptions may not be accurate in all cases (Beydoun et al., 2016; Emersic et al., 2015; Hiranuma et al., 2015), ns is commonly used to describe freezing data due, in part, to its simplicity. Here we use the following equation to calculate ns as a function of size from our experimental data: nsT=INPs(T)Stot, where INPs(T) is the INP concentration at temperature T determined from Eq. (2) in a given size range, and Stot is the total surface area of all aerosol particles in the same size range. Since this equation considers the surface area of all aerosol particles, rather than the surface area of just the INPs, the calculated ns values correspond to the total atmospheric aerosols.

Shown in Fig. 6 are the measured ns values as a function of size determined with Eq. (3). To generate Fig. 6, the aerosol surface-area concentration (Fig. 4b) was first binned using the same bin widths as the MOUDI, resulting in the total aerosol surface-area concentration in each size bin (Fig. S4b). Following Eq. (3), the INP concentration (Fig. 3) was then divided by the total aerosol surface-area concentration (Fig. S4b), resulting in ns values as a function of size. Figure 6 shows that ns is dependent on the size, with the larger particles being more efficient at nucleating ice. For Amphitrite Point and Labrador Sea, at a freezing temperature of -25 ∘C, ns was approximately 2 orders of magnitude higher for 8 µm particles compared to 0.2 µm particles. The dependence of ns on size can be qualitatively explained by considering four different types of aerosol particles, each having progressively larger geometric mean diameters and larger ns values. For example, consider a mixture of (a) sulfate aerosols internally mixed with black carbon with a small ns and a small geometric mean diameter, (b) sea salt aerosols with a larger ns and a larger geometric mean diameter, (c) clay particles with a larger ns and a larger geometric mean diameter, and (d) biological particles from terrestrial sources with the largest ns and largest geometric mean diameter. The assumption of a small ns for black carbon internally mixed with sulfate aerosols is consistent with previous measurements (e.g. Brooks et al., 2014; Chen et al., 2018).

To determine whether sea spray aerosols or mineral dust are the major sources of INPs at the three sites, Amphitrite Point, Labrador Sea, and Lancaster Sound, the measured ns values were compared to the ns values of sea spray aerosols and mineral dust at -15, -20, and -25 ∘C, respectively (Fig. 7). The ns values of sea spray aerosols in Fig. 7 are from field studies in the marine boundary layer and laboratory studies of sea spray aerosols as reported in DeMott et al. (2016). Specifically, the data in Fig. 1a in DeMott et al. (2016) were replotted and fitted using linear regression (Fig. S5). Since the reported ns values in DeMott et al. (2016) were based on dry, geometric diameters, they overestimate the ns values based on wet, aerodynamic diameters at 95 % RH by a factor of 6 (see Sect. S3). Figure 7 shows that the ns values of sea spray aerosols are smaller than the measured ns values in the supermicron range at all freezing temperatures at Amphitrite Point. This is also the case for Labrador Sea at freezing temperatures of -20 and -25 ∘C. For Lancaster Sound, the ns values of sea spray aerosols are smaller than the measured ns values for sizes of 5.6–10 µm and a freezing temperature of -25 ∘C. These combined results suggest that sea spray aerosols were not the major contributor to the supermicron INP population at Amphitrite Point and Labrador Sea and were not a major contributor to the largest INPs (5.6–10 µm in size) observed at Lancaster Sound.

The ns values of mineral dust particles shown in Fig. 7 are based on laboratory measurements with five different dust samples: Asian dust, Saharan dust, Canary Island dust, Israeli dust, and Arizona test dust (Niemand et al., 2012). Specifically, the data in Fig. 6 in Niemand et al. (2012) were replotted and fitted using linear regression (Fig. S6). Since the reported ns values in Niemand et al. (2012) were based on geometric diameters, they overestimate the ns values based on aerodynamic diameters by a factor of 2 (see Sect. S3). Figure 7 shows that the ns values for mineral dust are greater than or equal to the measured ns values at all three sites. These results suggest that mineral dust could be a possible source of the supermicron INPs at the three sites studied. However, these results do not confirm mineral dust as a major contributor to supermicron INPs nor do they rule out other types of particles as major contributors to supermicron INPs. Note that the data from Niemand et al. (2012) correspond to the ns values of mineral dust particles only, whereas the ns values reported here correspond to the total aerosol particles, as mentioned above. If we assume mineral dust particles are the only INPs in the atmosphere, and they account for 50 % of the total aerosol surface area, then the ns values of mineral dust shown in Fig. 7 divided by a factor of 2 would correspond to the ns values of the total atmospheric aerosols.

Shown in Fig. 8 is a comparison between the measured total INP concentrations (sum of the INP concentrations for all sizes measured) and the simulated INP concentrations at the surface at the three sites using a global model of INP concentrations based on the ice nucleation of K-feldspar and marine organics. When considering only marine organics as INPs in the model, predicted INP concentrations are less than measured INP concentrations in all cases except for Amphitrite Point at a freezing temperature of -25 ∘C. This suggests that sea spray aerosols are not the dominant source of INPs at the three coastal marine sites studied for all three temperatures, which is consistent with conclusions reached in Sect. 3.5. When considering only K-feldspar, associated with desert dust, as INPs in the model, the predicted INP concentrations at -25 ∘C are consistent with the measurements at all three sites, but at -15 and -20 ∘C, the predicted INP concentrations are less than measured. When considering both marine organics and K-feldspar as INPs in the model, the predicted INP concentrations at -25 ∘C are consistent with measurements, but at warmer temperatures, the predicted INPs are still less than measured. The underestimation of INP concentrations at warmer temperatures of the model could be explained by a missing source of INPs that are active at temperatures warmer than -25 ∘C, as hypothesized in Vergara-Temprado et al. (2017) based on the comparison with measurements at other sites. Possible sources missing in the model that could explain the high-temperature INPs include bacteria, fungal material, agricultural dust, or biological nanoscale fragments attached to mineral dust particles (Fröhlich-Nowoisky et al., 2015; Garcia et al., 2012; Haga et al., 2013; Mason et al., 2015a; Möhler et al., 2008; Morris et al., 2004, 2013; O'Sullivan et al., 2014, 2015, 2016; Spracklen and Heald, 2014; Tobo et al., 2013, 2014).

Recently Mason et al. (2015a) investigated the source of INPs at Amphitrite Point using correlations between INP number concentrations, atmospheric particles, and meteorological conditions. Correlations between INP number concentrations and marine aerosols (sodium as a tracer) and marine biological activities (methanesulfonic acid as a tracer) were not statistically significant. On the other hand, a strong correlation was observed between INP concentrations and fluorescent bio-particles, suggesting that biological particles from terrestrial sources were likely a dominant source of INPs at this site. These results are consistent and complementary to the studies presented above.

As discussed in Sect. 2.1.1, particle rebound from the substrate can be an issue when sampling particles with an inertia impactor. Good agreement between INP concentrations measured by the MOUDI-DFT and INP concentrations measured by a continuous flow diffusion chamber (a technique that is not susceptible to rebound) has been observed in previous field campaigns when the RH of the sampled aerosol stream was as low as 40–45 % (DeMott et al., 2017; Mason et al., 2015b). Nevertheless, particle rebound cannot be completely ruled out in the current study. If particle rebound was a factor when collecting particles with the MOUDI in the current study, the measured INP concentrations would be lower limits to the true INP concentrations, and the differences between simulated INP concentrations and measured INP concentrations shown in Fig. 8 would only be larger.