The role of nanoparticles in Arctic cloud formation

To constrain uncertainties in radiative forcings associated with aerosol–cloud interactions, improved understanding of Arctic cloud formation is required, yet long-term measurements of the relevant cloud and aerosol properties remain sparse. We present the first long-term study of cloud residuals, i.e. particles that were involved in cloud formation, and ambient aerosol particles in Arctic low-level clouds measured at Zeppelin Observatory, Svalbard. A detailed evaluation of the ground-based counter-flow virtual impactor inlet system is also presented. Cloud residuals as small as 15 nm are routinely observed especially 5 during the dark period and are potentially linked to ice, supporting prior work suggesting that classical droplet activation is not the only relevant process in the formation of Arctic low-level clouds. The reported measurements and findings provide a new basis for improving our understanding of Arctic clouds and for developing robust parameterisations of mixed-phase clouds in Earth system models.

Long-term observations of Arctic aerosol particles generally come from a relatively small number of permanent measurement stations. While there are differences in aerosol properties between the sites, it has been shown that they all share common 25 features both in terms of particle number concentration and particle number size distribution (Freud et al., 2017). This characteristic seasonal cycle of Arctic aerosol properties has been demonstrated previously for individual sites (Ström et al., 2003;Tunved et al., 2013;Nguyen et al., 2016). During the transition from winter to springtime, the number concentration of accumulation mode particles (typically diameter > 60 nm) increases due to long-range transport of polluted air masses -a phenomenon known as Arctic haze (Mitchell, 1956). In summer, changes in circulation and cloud cover lead to efficient scavenging of these 30 particles, subsequently lowering their concentration (Tunved et al., 2013). Lower accumulation mode particle concentrations, together with increased biological activity and photochemistry, helps facilitate new particle formation leading to number size distributions dominated by the smaller, Aitken mode particles (typically diameter < 60 nm) in the Arctic summertime (Ström et al., 2003). During autumn, the particle sinks are stronger than the sources because neither transport nor new particle formation is efficient, which leads to low number concentrations across the particle size spectrum (Tunved et al., 2013). 35 Studies characterising Arctic cloud condensation nuclei (CCN) generally cover short time periods, and only a couple of studies exist that look at the seasonal cycle in the Arctic (Jung et al., 2018;Dall'Osto et al., 2017;Schmale et al., 2018). Jung et al. (2018) measured CCN on Svalbard and found that the seasonal variation in CCN concentrations correlated well with the variation in accumulation mode aerosol particle concentrations. They also identified new particle formation and subsequent particle growth as contributors to summertime CCN concentrations, in line with results from a previous long-term study (Dall'Osto 40 et al., 2017) as well as shorter airborne and ground-based measurement campaigns (Leaitch et al., 2016;Zábori et al., 2015).
CCN number concentrations in the Arctic have been found to range between a few tens and a couple of hundred particles cm −3 (Jung et al., 2018), although concentrations vary spatially. Local concentrations of less than 1 and more than 1000 cm −3 have been reported (Mauritsen et al., 2011;Moore et al., 2011). CCN are of course only part of the picture -in cold and mixed-phase clouds, ice nucleating particles (INP) are also important. INP are much rarer, with concentrations several orders of magnitude and the ground-based counterflow virtual impactor (GCVI) inlet (blue) are connected to the differential mobility analysers (DMAs) and condensation particle counters (CPCs). The 3-way valve switches the sample flow to the instruments on the left-hand side from the GCVI inlet to the whole-air inlet when there is no cloud to be sampled. Cloud sampling is activated if the visibility drops below 1 km (measured by a visibility sensor (not pictured) next to the GCVI inlet). Auxiliary measurements from a fog monitor and an ultrasonic anemometer have also been included in the data analysis. and ambient cloud particle size distributions, meteorological parameters as well as remote sensing data, which, taken together, provide valuable new information about the elusive Arctic CCN. 60 2 Methods We present total particle and cloud residual size distributions and integrated number concentrations measured during more than 2 years (26 November 2015 to 4 February 2018) at Zeppelin Observatory using two different inlet systems. These measurements are complemented by measurements of ambient cloud particle size distributions, temperature, wind parameters and remote sensing data which are described below. A schematic illustration of the experimental set-up and a photo of the inlet systems the transmission efficiency of the inlet. Because the transmission efficiency depends on the size of the cloud particles before they are dried, it cannot be corrected for. However, an estimate of the absolute cloud residual concentrations can be obtained by back-calculating from the ambient cloud particle size distribution (as measured by a fog monitor) using the experimentally determined cloud particle size dependent transmission efficiency (Shingler et al., 2012) of the GCVI inlet. As will be shown below, we find that on average half of the ambient cloud particles make it into the GCVI sample flow, and cloud residual concentrations therefore have to be multiplied by a factor of 2 (see Sect. 3).  Table 1 for a summary of the instruments used, parameters measured and their temporal and/or spatial resolutions. We have not applied any standard temperature and pressure normalisation or particle shape correction to the data presented here, but multiple-charge corrections have been applied to all measured size distributions. They have also been corrected for particle losses due to diffusion and impaction using the Particle Loss Calculator by von der Weiden et al. (2009) Figure S2 shows how the DMPS systems compare during non-cloud periods. The comparison is based on data collected from May 2017-February 2018 (after the installation of the three-way inlet valve, see above). In general, the instruments compare well for large particle sizes while DMPS 2a-b shows consistently higher concentrations of small particles below around 30 nm in diameter. This is to be expected, since the diffusion losses are higher for DMPS 1 due to the instrument dimensions and longer sampling lines. These diffusion losses cannot be corrected for when the concentrations in the smaller size bins are zero.

175
Most of the differences originate from the lowest size bins between 10 and 15 nm, as can be seen in the scatter plots of Fig. S2 (panel c and d), where the integrated number concentrations of both DMPS 1 and DMPS 2a-b are shown. The slope of the orthogonal linear regression and the R 2 -value improve from 1.36 to 1.01 and 0.96 to 0.99, respectively, if particle number size distributions are integrated above 15 nm instead of 10 nm particle diameter.

180
A fog monitor (Droplet Measurement Technologies Inc., USA, Model FM-120) was used to determine the ambient cloud particle size and number concentration. It uses an optical method to size individual cloud particles at a flow rate of approximately 1000 L min −1 (airspeed 12 ms −1 ). The instrument is positioned facing south and measures ambient cloud particle size distributions in the size range 3.5-46 µm optical diameter (bin midpoints). More details on the instrument at Zeppelin Observatory can be found in Koike et al. (2019). It should be noted that no loss correction has been applied to the fog monitor data because 185 no clear signatures of particle loss were found in previous work by Koike et al. (2019), although significant sampling losses were suggested in other studies depending on, for example, the cloud particle diameter, and the wind speed and wind direction relative to the fog monitor (Spiegel et al., 2012).

Ultrasonic anemometer
A uSonic-3 Omni (METEK GmbH) ultrasonic anemometer was used to monitor wind conditions at Zeppelin Observatory. The

190
anemometer has 3 pairs of ultrasonic transducers arranged to form 3 paths along which the speed of sound is measured. From the difference in the travel time of sound along the 3 measuring paths, the 3D wind vector as well as the acoustic temperature can be derived. The acoustic temperature is a close approximation of the virtual temperature, which depends on the ambient relative humidity and is generally 1-2 degrees higher than the true temperature (METEK GmbH, 2013). However, note that at Zeppelin Observatory during our measurement period, the median difference between the measured acoustic temperature 195 and the ambient temperature measured by a Vaisala temperature probe (located in the meteorology mast at 15 m above the measurement platform) was around 3.4 • C. Nevertheless, we chose to use the temperature from our anemometer because we needed the higher time resolution that it provides.

Cloud remote sensing
The Cloudnet algorithm suite (Illingworth et al., 2007) has been applied to the Ny-Ålesund ground-based remote sensing 200 observations from the French-German research station AWIPEV (Nomokonova et al., 2019b), which is located approximately 2 km north of Zeppelin Observatory. A standard product is the target classification which combines measurements from cloud radar, ceilometer and microwave radiometer with output from a numerical weather prediction model. Each radar height bin is classified in terms of the occurrence of e.g. liquid droplets, ice particles, rain/drizzle, melting ice and a combination of those.
More details on the product for Ny-Ålesund can be found in Nomokonova et al. (2019b). For comparison with the cloud 205 residual data collected at Zeppelin Observatory, we selected Cloudnet height bins between 400 and 600 m. We only compared cases when the cloud base height at AWIPEV was between 300 m and 600 m to ensure that the classifications were likely to be applicable also to the cloud at Zeppelin Observatory. It should be noted that this cloud base height criterion reduces the number of data points we can use, such that we only have Cloudnet data for approximately 30% of our in-cloud size distribution data.

210
A cluster analysis was performed to identify cloud residual size distributions that were dominated by Aitken mode particles.
We used k-means clustering, implemented in the scikit-learn (v. 0.20.2) Python package (Pedregosa et al., 2011), which is a method to categorise data into a pre-defined number of clusters, k, where members of a cluster are as similar to each other as possible while at the same time being as different to members of other clusters as possible. Each data point is assigned to the cluster with the nearest mean. We categorised cloud residual number size distributions based on their shape, so the size 215 distributions were normalised by the integral before applying the k-means algorithm. We selected 5 clusters (k = 5) to separate out the cloud residual size distributions that were dominated by the very smallest particles. Choosing fewer clusters did not fully separate this distribution of interest, while more clusters led to a further splitting of the accumulation mode (see Fig. S3).
The ambient cloud particle size distributions were used to evaluate the sampling performance of our GCVI system. Assuming 220 that the cloud particle distribution measured by the FM-120 fog monitor is an accurate representation of the cloud particles that enter the GCVI inlet, we applied the experimentally determined size-dependent transmission efficiency from Shingler et al.
(2012) (linearly extrapolated to cover the full FM-120 cloud particle size range) to calculate the cloud particle concentration above the GCVI cut-size that would have made it into the sample flow. Here, it is important to note that the transmission efficiency was determined for hollow glass beads without using the inlet counterflow (Shingler et al., 2012). As such, it does 225 not take into account potential evaporation of water from the cloud particles in the different inlet segments. Within this work, we have only used the transmission efficiency determined for the first inlet segment, because we believe that the dry counterflow initiates evaporation which would make the transmission efficiency determined for subsequent sections an underestimation of the true transmission efficiency. This choice may result in an overestimation of the transmission efficiency (particularly of larger cloud droplets) since some losses are effectively ignored, but no correction is preferable to an invalid correction.

230
The corrected cloud particle concentrations, integrated above the GCVI cut-size, were compared to the integrated cloud residual number concentrations measured behind the GCVI inlet, and the result can be seen in Fig. 2. Given the uncertainties involved, the instruments agree reasonably well in terms of the seasonal cycle and magnitude of cloud particle/cloud residual concentrations (Fig. 2a). A 2D histogram of corrected cloud particle concentrations versus cloud residual concentrations ( Fig. 2b) shows that most of the data points lie on or around the 1:1 line. An orthogonal linear regression of cloud residual ver-235 sus cloud particle number concentrations (Fig. 2b) returns a slope of 0.97, an offset of 4.9 cm −3 and an R 2 of 0.47. However, there is a substantial amount of scatter. Most notably, there is a cloud of data points below the 1:10 line (∼ 7-8% of the data) that seems to be associated with colder temperatures at the sampling site ( Fig. 2c).
Temperatures below 0 • C could indicate the presence of ice crystals (e.g. in ice or mixed-phase clouds). Both the GCVI and the FM-120 were calibrated using spherical particles, which makes the comparison especially difficult for cases when ice 240 crystals are sampled. The true transmission efficiency of the GCVI inlet is going to be different for non-spherical particles, i.e. ice crystals, which are not accurately represented by glass beads. In addition, the concept of size becomes ambiguous when the sampled particles are not spherical, especially since the two instruments deal with different types of size. The optical size reported by the FM-120 is not necessarily the same as the Stokes equivalent size that determines how a crystal behaves inside the GCVI inlet, which means that the transmission efficiency we apply could be incorrect. For non-spherical ice crystals, the 245 sizing uncertainties in the fog monitor can be larger than those associated with Mie theory (Baumgardner et al., 2017), and as a result under-or oversizing of crystals can occur. This could also affect the concentration comparison. However, the points below the 1:10 line in Fig. 2 (∼ 7-8% of the data) still remain below the 1:1 line even if we compare the cloud residual concentration to the total, uncorrected cloud particle concentration (not shown), which suggests that something other than errors in the assumed transmission efficiency is causing the difference.

250
Riming or impaction scavenging of interstitial aerosol particles onto an ice crystal could result in more than one cloud residual emerging from the crystal as it dries inside the GCVI inlet (e.g., Mertes et al., 2007;Santachiara et al., 2018). Similarly, an INP could break or eject material during the freezing process (e.g., Lauber et al., 2018) which could also result in more than one residual particle per crystal. These processes could make cloud residual concentrations exceed cloud particle concentrations; however, the concentrations in Fig. 2 sometimes differ by almost two orders of magnitude, and it is unlikely that the 255 aforementioned processes could account for the full difference (Santachiara et al., 2018).
Measurement artefacts during in-situ sampling of cloud droplets and ice crystals are a common and complex challenge (Baumgardner et al., 2017) contributing to both overestimation and underestimation of cloud residual (or cloud particle) con-centration measurements (Pekour and Cziczo, 2011;Spiegel et al., 2012;Shingler et al., 2012). Particle capture by wake effects in the GCVI inlet is a possible explanation for cloud residual concentrations exceeding cloud particle concentrations; however, 260 even at cloud particle and interstitial aerosol particle concentrations at the upper end of our observed values, only 1 % of the measured cloud residuals is estimated to be a potential artefact (Pekour and Cziczo, 2011). Thus, this effect is likely not the major cause of the disparity between cloud residual and cloud particle concentrations that sometimes occurs. Droplet or ice crystal shattering is another potential source of small particles within the acceleration and deceleration zones of the GCVI, and this could also cause an overestimation of the cloud residual number concentration. Intuitively, one would expect shattering 265 events to produce large amounts of particles, yet the largest relative differences between cloud particle and cloud residual concentrations mainly occur at very low particle concentrations. Nevertheless, the apparent correlation with cold temperatures ( Fig. 2c) means that ice crystal shattering inside the GCVI inlet cannot be fully ruled out.
A comparison of the measured visibility and the visibility calculated from the FM-120 data shows a reasonable agreement for the majority of data points, but again there is a group of data points at predominantly cold temperatures where the agreement We have already suggested that the assumption of spherical particles might not hold at cold temperatures, which could explain the differences in Fig. S4. However, the calculated visibility is sometimes several orders of 275 magnitude higher than the measured one, and it seems unlikely that non-sphericity would cause such large differences. The presence of precipitating particles may cause the measured visibility to be higher than the calculated one; however, precipitating particle concentrations at Zeppelin Observatory have previously been found to be mostly lower than 0.3 cm −3 . Hence, while the presence of such particles could explain the differences in visibility, they do not explain the differences in concentration in Fig. 2. Local effects of blowing snow could also affect the measured visibility, but there is no 280 apparent correlation between wind speed and differences in cloud particle and cloud residual concentrations (see Fig. S5 in supplementary material). Thus, we are left with the possibility that the differences in visibility are caused by a loss of detected cloud particles within the FM-120. It could be that ice crystals are more susceptible to losses due to turbulent deposition inside the contraction of the inlet (Spiegel et al., 2012).
Recalculating the visibility after scaling up the cloud particle size distributions to the measured cloud residual concentrations 285 (i.e. multiplying each cloud particle size distribution by the ratio of the total cloud residual concentration and the cloud particle concentration integrated above the GCVI cut-size) significantly improves the agreement with the measured visibility (see Fig. S4c-d in supplementary material). Note that the concentrations were only scaled up, not down, since the total cloud particle concentrations could be higher than the cloud residual concentrations due to the cut-size of the GCVI inlet. The improvement in the visibility comparison after scaling the cloud particle concentrations points towards an undercounting of 290 cloud particles in the FM-120 for parts of the period with colder temperatures. Thus, it cannot be ruled out that the discrepancy between the measured and calculated visibilities, and by extension the discrepancy between cloud residual and cloud particle concentrations, is partly due to the FM-120 undercounting the ambient ice crystals.
As we have seen, interpreting cloud residual data is a non-trivial task. There are many processes, both natural and instrument related, that can influence the measured concentrations of both cloud residuals and cloud particles. We cannot definitively 295 say which sources of error are affecting which instrument, and therefore we will not discard any data at this stage. Still, a more detailed comparison with respect to temperature and wind parameters -also including data from the whole-air inlet -is warranted, and this is done in the next subsections.
In Fig. 2, we corrected the cloud particle concentrations for the GCVI transmission efficiency. However, to be able to compare our cloud residual measurements to the aerosol particle measurements from the whole-air inlet, we need to apply the correction 300 in the other direction. The integrated transmission efficiency of the GCVI inlet was estimated by calculating the ratio of the integrated cloud particle number concentrations with and without taking into account the size-dependent transmission efficiency of Shingler et al. (2012). Figure S6 shows the results when integrating over the entire cloud particle population and when integrating only above the cut-size diameter of the GCVI. Above the GCVI droplet/crystal cut-size (red histogram in Fig. S6a), the distribution is symmetrical and relatively narrow, and it shows that approximately half (mean ± std ratio of 0.5 305 ± 0.05) of the total cloud particles were sampled. The mode of the distribution shows very little variation between seasons ( Fig. S6b). We have therefore corrected all cloud residual size distributions and concentrations by a factor of 2 assuming cloud residual size and cloud particle size are not correlated. An individual correction factor for each data point would, in theory, be possible, but only for the points where we have overlapping cloud particle and cloud residual data. Thus, for the sake of consistency, we use a constant correction factor. 3.1 Comparison with respect to ambient (acoustic) temperature Figure 3 shows concurrent cloud particle, cloud residual and total aerosol particle data binned by ambient temperature (acoustic temperature recorded by the anemometer). Note that the acoustic temperature was, on average, 3.4 • C higher than the actual air temperature (see Sect. 2.3.3) at Zeppelin Observatory during our measurement period. Panel a shows box plots of cloud residual concentrations (corrected by the factor 2, see above) and cloud particle concentrations (now without correction, but 315 still integrated only above the GCVI cut-size). The concentrations agree well down to about -2 to -4 • C, where the cloud particle concentrations drop below the cloud residual concentrations. These bins contain relatively few data points (bar plot in Fig. 3a), but they follow the general trend of decreasing cloud particle/cloud residual concentrations with decreasing temperature. Figure 3b shows the mean cloud residual and total aerosol particle size distributions for the same temperature bins. Both the Aitken and the accumulation mode are present in both cloud residuals and total particles, but the total particle size distributions 320 generally show higher particle concentrations, particularly of Aitken mode particles. Figure 3c shows the ratio between the distributions in panel b, i.e. cloud residual concentrations divided by total particle concentrations. At temperatures above approximately -2 • C, the resulting curves are sigmoidal like typical CCN-activated particle fraction curves. During cloud events at these temperatures, the figure shows that most of the total aerosol particles larger ∼100 nm are in fact cloud residuals. The apparent D 50% , defined as the diameter where the ratio is around 0.5, decreases with decreasing temperature, which could 325 be related to the general decrease in particle concentrations with temperature seen in the other two panels of the figure, since lower overall particle concentrations could allow smaller particles to activate (assuming liquid droplet activation without size- At temperatures below -2 • C (approximately 16 % of this subset of data), however, the curves in Fig. 3c look very different.
Instead of an S-shape, the curves are relatively flatter with a maximum appearing at lower sizes, with the coldest temperature bins even showing a peak below ∼20 nm particle diameter. This implies that accumulation mode particles have not acted as 335 cloud residuals, while now an increased contribution of Aitken mode particles served as cloud seeds. These clouds most likely contain ice particles and the question arises if the small particles could potentially be caused by sampling artefacts inside the CVI sampling line (see Sect. 3 above) or if real physical atmospheric process is underlying this observation. While artefacts cannot be completely ruled out (see Sect. 3), it should be noted that approximately the same size modes are present in the whole-air inlet (Fig. S7), and there is no reason to expect that any potential droplet or crystal shattering in this inlet should 340 produce the same size particles as shattering in the GCVI inlet. The shape of the ratio curves in Fig. 3c below -2 • C indicate the participation of ice in cloud formation as will be discussed in Sect. 5 within a cluster analysis of the residual size distributions. . c Ratio of the size distributions in b, i.e. cloud residual concentrations divided by total particle concentrations.

Comparison with respect to updraft
The FM-120 as well as the GCVI inlet sampling efficiency can also be affected by the wind speed and direction (Spiegel et al., 2012), but this is not something we can easily correct for. However, heatmaps similar to Fig. 2c for wind speed, updraft, and 345 wind direction indicate no obvious correlation between wind parameters and deviations of concentrations from the 1:1 line (see Fig. S5). One should also take into account that high wind speeds are only rarely observed at Zeppelin Observatory (the median wind speed is approximately 3 ms −1 ), which can be seen in Fig. S1. Nevertheless, since Zeppelin Observatory is a mountain site, a closer look at the updraft is warranted to investigate potential orographic effects. Figure 4 shows concurrent cloud particle, cloud residual and total aerosol particle data, this time binned by updraft veloc-350 ity instead of acoustic temperature. The box plots in the first panel show that the cloud residual and cloud particle number concentrations generally agree well, but there seems to be a tendency for the cloud residual number concentrations to be underestimated at higher updraft, starting approximately above 1 ms −1 . Panels b and c show a similar pattern, with cloud residual concentrations decreasing in the last four updraft bins. This pattern is not observed in the total aerosol particles, except for in the highest updraft bin (Fig. 4b). The curves in Fig. 4c systematically level out at lower ratios with higher updrafts (for the last 355 four bins), which could either mean that not all accumulation mode particles are cloud residuals under these conditions, or that the GCVI inlet fails to sample all cloud particles at high updrafts. Taken together with the previous panels, it seems likely that the GCVI inlet sampling efficiency is negatively affected by high updraft velocities (or indeed high wind speeds in general, as these parameters tend to be correlated at Zeppelin Observatory). The sampling efficiency of the FM-120 fog monitor can in theory also be adversely affected by high wind speeds (Spiegel et al., 2012), but this seems to happen to a lesser extent than for the GCVI inlet based on Fig. 4a. One should bear in mind that the wind speeds and updrafts are generally lower near the FM-120 as it is positioned at a lower altitude than the GCVI inlet (∼5 m below).

The annual cycle of Arctic cloud residuals at Zeppelin Observatory
During cloud events within our measurement period, typical cloud residual number concentrations ranged between 10 and 62 cm −3 (25 th and 75 th percentiles), with a median of 25 cm −3 (mean ± standard deviation: 50 ± 66 cm −3 ). Total concentra-365 tions of particles suspended in the air (diameters 10-809 nm) during these cloud events were generally higher, ranging up to 163 cm −3 (75 th percentile) with a median of 70 cm −3 (mean ± standard deviation: 145 ± 235 cm −3 ).  The observed total particle number concentrations follow the typical seasonal cycle of Arctic aerosol. We recognise the characteristic maxima in number concentration due to Arctic haze in spring and new particle formation in summer, and the low, relatively stable concentrations during the rest of the year. There are some differences compared to previous measurements 375 at Zeppelin Observatory, for example in terms of when peak concentrations occur (e.g., Ström et al., 2003;Tunved et al., 2013;Freud et al., 2017). Such differences could be due to annual variability, which has previously been shown to be significant (Freud et al., 2017). In addition, it should also be kept in mind that we present number concentrations exclusively during cloud events and concentrations shown are for particles above 20 nm diameter, in contrast to previous studies.
The cloud residual number concentrations, while lower than the total particle concentrations, display a similar seasonal 380 behaviour. The shape and the magnitude of the cloud residual annual cycle are confirmed by ambient cloud particle measurements (cf. Fig. 2a). Based on Fig. 5a, the springtime peak in cloud residual concentrations appears to lag that of the total aerosol particles by one month -the maximum cloud residual concentration occurs in May rather than in April. However, this apparent shift does not appear if we only consider low updraft cases (Fig. 5b). The total particle concentration peak in April in Fig. 5a appears to be driven by the high total particle number concentrations in April during high updraft events (Fig. 5c). Most of 385 the April data points in Fig. 5c are from the same event -a relatively thin cloud where a large difference between the cloud residual and total particle concentrations was observed. Overall, the high updraft cases are characterised by very low cloud residual number concentrations and the shape of the annual cycle is slightly different from the other two panels. This may indicate that an explanation could be the potential decrease in the GCVI sampling efficiency at high updrafts (see Sect. 3.2).
One should also bear in mind that the months of March and April are characterised by a low number of observations (80-100 390 size distribution scans), limiting the statistical significance for these months for the concurrent data.
In terms of number size distribution (see Fig. 6), the cloud residual population is dominated by accumulation mode particles during most of the year. However, the Aitken mode is often also present, and there appears to be a clear seasonality in the relative abundance of Aitken and accumulation mode cloud residuals. In January and February, the median size distributions are dominated by the Aitken mode. In spring, particularly April and May, there are very few Aitken mode cloud residuals in 395 comparison to the number of accumulation mode residuals. The size distributions then become more bimodal during summer and autumn, then in December the Aitken mode dominates again. The size distributions in Fig. 6 are normalised to highlight their shape. Non-normalised monthly cloud residual size distributions, together with concurrent total aerosol size distributions, can be found in Fig. S8 in the supplementary material. Note, however, that the figures show different subsets of the data - Fig. 6 shows all cloud residual data we have, while Fig. S8 is limited by the availability of concurrent data from the whole-air inlet.

400
Generally speaking, the total aerosol size distributions and the cloud residual size distributions show similar size modes, but the total aerosol concentrations are higher, particularly for the Aitken mode.
In Fig. 4, clear changes in the size cloud residual distributions (i.e. shift towards smaller activation diameters) are only seen for updrafts above around 2 ms −1 , which is a relatively small subset of the observations. The overall shape of the monthly averaged cloud residual size distributions do not change significantly if only taking into account cases with updrafts below 405 1 ms −1 (see Fig. S9 in the supplementary material). As such, we will use all size distributions for all updrafts from now on.
5 The importance of Aitken mode particles in Arctic clouds Figure 6 shows that Aitken mode cloud residuals, even below 30 nm in diameter, occur throughout the year at Zeppelin Observatory. In this size range, particles are often not considered to be potential CCN (nor INP), but a closer look at our measured cloud residual size distributions shows that Aitken mode particles often make up a significant part of the total cloud residual 410 number concentration. Figure 7 shows the seasonality of the contribution of sub-100 nm, sub-50 nm and sub-25 nm particles to the overall measured cloud residual population. The sub-100 nm size range is included for illustrative purposes, since 100 nm is sometimes used as a lower size threshold for particles to be considered CCN-active for liquid clouds (Kuang et al., 2009;Yu et al., 2014;Patoulias et al., 2015). The fractions presented in Fig. 7 have been calculated based on daily mean cloud residual concentrations and show that particles smaller than 100 nm in diameter make up between 30 and 70 % of the total measured 415 cloud residual number concentration in the majority of the cases. In fact, the average contribution (both mean and median) is close to or above 50 % in all but four months. The months where the sub-100 nm cloud residuals make up a smaller fraction of the total are the months when the total aerosol particle number concentration is the highest (April through July; cf. Fig. 5).
The seasonal pattern looks similar for all three cloud residual size ranges in Fig. 7. The relative contribution of Aitken mode particles to the total cloud residual number concentration increases during autumn and continues to do so until it reaches a 420 maximum in February. Then, when the haze period starts in March and April, the contribution of Aitken mode particles begins to decrease as the number of accumulation mode particles increases. The relative contribution of sub-100 nm particles is at its lowest in April, while the sub-50 nm and sub-25 nm relative contributions continue to decrease until June. This behaviour is opposite to the total aerosol particles, where the summer months are the months with the highest relative contribution of Aitken mode particles due to the increased contribution of new particle formation (cf. Fig. S10).

425
To find out under which conditions the smaller residuals are present in the cloud particles, we performed k-means clustering on the cloud residual number size distributions (normalised by the corresponding total cloud residual number concentration).
The results when 5 clusters are used are presented in Fig. 8. The clusters are numbered from 1 to 5 according to increasing modal diameters of the cluster average size distributions (the approximate modal diameters are 15, 30, 65, 100, and 150 nm).
This order is also reflected in the total number concentration (cf. Fig. S11). Cluster 2 is the most frequent cluster (27 % of the 12 µm cloud particle diameter. The Aitken mode clusters, on the other hand, are associated with larger cloud particles (Fig. 8a) and the lowest cloud particle and cloud residual number concentrations (Fig. S11), suggesting they represent optically thin clouds with few, large droplets and/or ice crystals. This is further corroborated by the visibility distribution ( Fig. S11) which shows high values, in particular for Cluster 1. Cluster 1 also stands out in that it occurs primarily during the winter months and at low temperatures (Fig. S11). No clear relationship between these two clusters and wind speed or updraft was found.

440
The ratio between cluster mean cloud residual size distributions and corresponding mean ambient particle size distributions, for liquid clouds this would correspond to the activation ratio, are shown in Fig. S7 in the supplementary material. Cluster 1 and (to a lesser degree) Cluster 2 clearly deviate from the classical Köhler theory of droplet activation assuming a sizeindependent chemical composition. It is possible that differences in particle composition could explain part of this behaviour (e.g., McFiggans et al., 2006;Lowe et al., 2019); however, liquid droplet activation is not the only relevant process at our site 445 and therefore the cloud residual distributions we measure could also be related to ice processes. The fact that Cluster 1 occurs predominantly during winter and at lower temperatures than the other clusters (Fig. S11) would also be consistent with an influence from ice processes.
Comparing the target classification of Cloudnet above Ny-Ålesund, at the altitude around Zeppelin Observatory, to the cluster analysis indeed shows a higher occurrence of ice for Cluster 1 ( Fig. 9  by category 1) decreases from Cluster 1 to Cluster 5, which is consistent with the activation ratios in Fig. S7 which appear more like classical Köhler activation (of homogeneously mixed particles) when moving from Cluster 1 to Cluster 5. However, it should be noted that all Cloudnet classification categories appear in each of the clusters (Fig. S12), so the cloud residual size

Discussion
Results presented in this paper are the first direct long-term measurements of size resolved cloud residual number concentrations of low-level clouds in the Arctic. It is also the first cloud residual dataset that covers more than a full annual cycle, in the Arctic and globally. It includes the important winter months, when Arctic warming is most pronounced (Maturilli and Kayser,460 2017) and clouds are hypothesised to play a key role. Our measured cloud residual number concentrations generally follow the typical annual aerosol cycle previously reported for this site (Tunved et al., 2013). During the autumn and winter months, we found, in relative terms, a significant contribution of Aitken mode particles to the cloud residual number concentration.
Cloud residual measurements differ from standard CCN measurements in that instead of attempting to replicate in-cloud conditions inside the instrument -most notably fixed supersaturation bands in place of dynamic ambient conditions -we 465 extract cloud particles from the air, dry them and subsequently count and size the cloud residuals. While there are only a few long-term datasets from the Arctic (Jung et al., 2018;Dall'Osto et al., 2017;Schmale et al., 2018) that we can compare to, this difference in measurement techniques seems to be important. Jung et al. (2018) found that CCN concentrations correlated well with concentrations of accumulation mode particles at Zeppelin Observatory, and that median CCN concentrations peaked in March at most supersaturation levels. This is different from our measured cloud residual concentrations, which peak in May.

470
However, one should keep in mind that Jung et al. (2018Jung et al. ( ) considered different years (2007Jung et al. ( -2013 and did not differentiate between in-or out-of-cloud periods. In addition, Jung et al. (2018) observed for most of the year higher CCN concentrations than our cloud residual concentrations, particularly in winter, highlighting the differences between measurement techniques.
Studies where particles are artificially activated, i.e. at a fixed supersaturation, are independent of the ambient meteorology and atmospheric dynamics, whereas our study inherently takes the ambient conditions into account by sampling the actual 475 ambient cloud droplets or ice crystals. Therefore, the differences between our observed concentrations could either be because the actual ambient supersaturations are lower than what is used in CCN counters, or because ice processes are involved while CCN counters only consider liquid droplet activation.
The importance of Aitken mode particles for Arctic clouds has previously been shown (e.g., Leaitch et al., 2016;Koike et al., 2019;Korhonen et al., 2008), however, by indirect means or model studies. Our results support these findings with direct 480 measurements of cloud residuals. Furthermore, we find that Aitken mode particles also play an important role in wintertime clouds at Zeppelin Observatory, while previous observations have focused on the Arctic summer months. The clear seasonality we observe in the relative contribution of small particles to the cloud residual number concentration could partly be explained by the interplay between aerosol particle sources, sinks, meteorology and condensible water vapour. In late autumn and winter, aerosol particle concentrations decrease rapidly (Fig. 5) and the Arctic atmosphere becomes drier (Maturilli and Kayser, 2017).

485
However, if the decrease in condensation sink, due to reduced particle concentration, is larger than the decrease in water vapour, there will be, in relative terms, more water vapour available for fewer particles in winter. These conditions allow for higher supersaturation to be reached and smaller particles to be activated (assuming no strong seasonal cycle in the updraft velocity as observed here, see Fig. S1). In other words, the winter season at Zeppelin Observatory falls into the CCN-limited cloudaerosol regime that has previously been reported for the summertime High Arctic (Mauritsen et al., 2011;Leaitch et al., 2016).

490
However, this only applies to liquid clouds, while the cloud residuals we measure could correspond to either CCN or INP.
Unfortunately, no cloud phase data are available for our measurement period but, by proxy of the Cloudnet target classification from above Ny-Ålesund, we have shown that the cloud residual size distributions dominated by the very smallest particles are likely to be influenced by ice processes.
Some of the cloud residuals we have measured, in particular those in Clusters 1 and 2, are much smaller than typical INP 495 (Hoose and Möhler, 2012;DeMott et al., 2010) (or indeed CCN). Yet residual size distributions with a similar shape have previously been observed for ice particles in mixed-phase clouds measured with an Ice-CVI (Mertes et al., 2007), and droplet residuals down to 25 nm diameter have previously been observed by GCVI measurements (Schwarzenboeck et al., 2000) and predicted in model studies (Gérémy et al., 2000;Korhonen et al., 2008).
A question that arises is where these small particles come from. In the Arctic and marine boundary layer, the presence 500 of particles below ∼50 nm is most often associated with new particle formation (Ström et al., 2003;Tunved et al., 2013) or primary emissions of sea spray particles (Ovadnevaite et al., 2011). However, these sources are unlikely to explain the presence of small particles during winter, when there is reduced or no sunlight (i.e. no photochemistry), less biological production and most of the sea surface is covered by ice (Dall'Osto et al., 2017;Sharma et al., 2012). Other potential sources can be long-range transport, but the lifetime of Aitken mode aerosol particles in the boundary layer is rather limited, or entrainment from the free 505 troposphere. However, this is purely speculative and future studies are needed to investigate the exact sources and chemical nature of these small particles.
As the remote-sensing results suggest, the ice phase appears at the height of Zeppelin Observatory predominantly during mixed-phase cloud conditions. It could potentially be that the crystals we measure are the result of secondary ice formation processes (Field et al., 2016), which has been suggested in a model study to be important for Arctic stratocumulus clouds 510 (Sotiropoulou et al., 2020). This could include a distribution of the original CCN or INP material to the ice splinters, which act as new nuclei to further ice particle formation. In other words, the cloud residuals we measure do not have to correspond to single CCN or INP but may also be fragments of these which would explain their small size. The shape of the cloud residual size distribution of Cluster 1 compared to the ambient particle size distribution (cf. Fig. S7a) reveals that the accumulation mode particles do not activate. This points towards a possible water vapour transfer to the ice splinters via the Bergeron-Findeisen 515 process causing a larger concentration of interstitial aerosol particles. However, it should be noted that secondary ice particles cannot be the only reason for the small residuals we observe, as it would not explain why the cloud residual and cloud particle concentrations do not always agree during these cases (see e.g. Fig.3a) (unless the ice crystals are undersampled by the fog monitor, see Sect. 3). There are other processes, such as riming, which could be consistent with both small residuals and a discrepancy between cloud residual and cloud particle concentrations. Additionally, since mixed-phase clouds are concerned, 520 part of the cloud residuals will of course also come from liquid cloud droplets. It is not possible to tell if a given cloud residual is a result of liquid droplet activation (CCN), ice nucleation (INP), or secondary processes (CCN/INP fragment) without further detailed information on cloud phase, structure and origin. The cloud phase is an important parameter should as such be added in future studies.
Although an overall good agreement between ambient cloud particle and cloud residual number concentrations is found, one 525 has to keep in mind that measurement artefacts can still not be fully excluded given the complexity of our observations. The initial data set was carefully screened for malfunctioning of instrumentation and local contamination. A thorough assessment of potential artefacts and instrument uncertainties was made in Sect. 3. While there were some cases where the agreement between the GCVI and the FM-120 was clearly worse, we did not discard these data because we were unable to prove that the disagreement was only caused by artefacts in the GCVI. In addition, the disagreement is not completely random -data points 530 at cold temperatures are overrepresented (see Figs. 2 and S4). Thus, removing these points would have introduced a (potentially unjustified) bias into our analysis. Despite the uncertainties, we believe that our results are reliable enough to show that small particles are likely contributing to the formation of mixed-phased clouds at Zeppelin Observatory. This is especially important during the dark period when overall aerosol concentrations are low and even small changes in available CCN and INP will have a strong impact on cloud properties.

Conclusions
Our study presents a unique seasonal picture of aerosol particle activation in clouds in a polar environment and is the first longterm study of cloud residuals in the Arctic. Activation of aerosol particles in low-level clouds in the Arctic is strongly coupled to the annual variability of aerosol particle number and size distribution, meteorology and the availability of water vapour. We have demonstrated that the use of CCN proxies with fixed size limits (e.g. 100 nm diameter) or accumulation mode aerosol 540 particles is incorrect for the Arctic environment, where smaller particles act as CCN, supporting the results of previous studies.
For large parts of the year, and especially during the dark period, we observed a large relative contribution of Aitken mode particles to the cloud residuals, which could be explained either by a strong CCN-limited regime or ice processes, possibly including secondary ice formation. In winter, the Arctic exhibits the strongest warming trend and clouds are believed to play an important role in this process. Even subtle changes in aerosol particle number concentrations during the dark period in the Arctic can result in large effects on cloud microphyscial properties and thus also perturb cloud-related warming effects by changing the radiation balance in the infrared spectrum. The climatology presented here provides a new benchmark dataset for further model-measurement evaluation exercises to improve the representation of low-level clouds in Earth system models.
Our work also shows the importance of focusing more research in the Arctic on the dark period. We have demonstrated the experimental complexity involved in aerosol-cloud interaction research, highlighting the strengths and weaknesses of sampling 550 cloud droplets and crystals by means of the GCVI technique. The direct measurements of cloud residuals provide a valuable new perspective on Arctic CCN and INP, but information about the cloud particle phase and the residual chemical composition would be necessary to be able to disentangle and better understand all the processes and particle sources involved in Arctic cloud formation.
Data availability. The data of this study will be available on the Bolin Centre Database (DOI and link will be added later). The Cloudnet