The number concentration of INPs ([INP]) was determined using the Horizontal Ice Nucleation Chamber (HINC; Lacher et al., 2017) at -31 ∘C (±0.4 ∘C; [INP]-31) and at a relative humidity with respect to water (RHw) of 103 % (±2 %), representing condensation and/or immersion freezing, relevant for the formation of MPCs (e.g., DeBoer et al., 2011; Murray et al., 2012; DeMott et al., 2015, 2018). A nucleation temperature of -31 ∘C was chosen in order to avoid measurements below the limit of detection (LOD) of the instrument. This is crucial at a remote location such as JFJ where the [INP] is naturally low but is generally higher at colder nucleation temperatures. Moreover, a nucleation temperature of -31 ∘C ensures that a large spectrum of INP activation at MPC conditions is covered. The RHw of 103 % ensures that the entire aerosol layer, which experiences a varying RH between 101 %–103 %, is above water saturation such that the particles can activate into droplets in the given residence time of HINC. Cloud-like temperature and saturation conditions in HINC are established by applying a temperature gradient along ice-coated parallel plates, held at sub-zero temperatures. INPs are sampled into the chamber guided within a sheath flow and exposed to the temperature and supersaturation in the center of the chamber and can activate into water droplets and/or ice crystals. Due to particle losses in the system, mainly due to the tubing upstream of the chamber and its horizontal orientation, the measurements are representative for aerosol particles <2 µm (56 % transmission efficiency at this size), which is the dominant size range for the aerosols present at the site (Baltensperger et al., 1997; Nyeki et al., 1998). We note that in this way some of the larger (>2 µm) particles may not be detected. For more details on the transmission efficiency see Lacher et al. (2017). [INP]-31 was measured with a time resolution of 20 min, which is alternated with 10 min of background measurements via a filter to exclude the impact of counting undesired frost particles generated within HINC. The instrument's LOD is thereby calculated based on these background measurements. To achieve better measurement statistics due to the naturally low [INP] in the free troposphere, an aerosol particle concentrator (the portable fine particle concentrator, PFPC; Gute et al., 2019) was deployed upstream of HINC during the field campaign, allowing an enrichment in aerosol particles >0.1 µm. The enrichment is thereby size dependent due to the working principle of the PFPC, with an enrichment factor of ∼10 at particle sizes of 0.3 µm and a maximum enrichment of ∼20 for particles >0.75 µm (Gute et al., 2019). The INP enrichment factors were determined by consecutive measurements on and off the concentrator and showed a large variability between values of 1 and 23, reflecting the variability in the size of the present INP population (see Lacher et al., 2018a, for a more detailed description of this setup). For comparison and correlation with other measurements, only quantifiable [INP]-31 > LOD (∼0.2stdL-1) was considered for this study. Here we present results from [INP]-31 and its related ice-active surface site density (ns), which normalizes [INP] to the available surface area per volume of air of the ambient aerosol particles (Connolly et al., 2009; Hoose and Möhler, 2012): 1ns=[INP](#m-3)total particle surface area(m2m-3) given in m-2, and which requires the surface area distribution concentration to be calculated from the number size distribution. Measurements from the GAW SMPS and OPS are used to calculate the available surface area of the aerosol population by assuming a uniform particle shape and that the refractive index of the ambient aerosol population is represented by the calibrated value of the OPS. We acknowledge that this can lead to higher uncertainties, which are not quantified here. The concept of ns is based on the assumption of a uniform composition of the investigated aerosol sample and assumes that the temperature dependence of ice nucleation is more important than the time dependence, which therefore can be neglected (e.g., Welti et al., 2012).

In order to investigate the relationship between [INP]-31 and ns with meteorological and aerosol parameters, we use Spearman's rank correlation coefficient (Spearman, 1904). Spearman's rank correlation determines to what extent two variables are monotonically related by applying a linear correlation analysis to the rank-ordered values of the parameters. As we would not necessarily await a linear relationship amongst parameters in atmospheric science, this test is well suited for our purposes. Examples for the correlation analysis are presented in Fig. S6 in the Supplement.

Filter-based aerosol collection followed by offline quantification of [INP] was performed with the FRankfurt Ice nucleation Deposition freezinG Experiment (FRIDGE; Bundke et al., 2008; Klein et al., 2010; Schrod et al., 2016) and the Ice Nucleation Spectrometer of the Karlsruhe Institute of Technology (INSEKT), which is a re-built version of the ice spectrometer freezing method (Garcia et al., 2012; Hill et al., 2016).

For FRIDGE, the aerosol particles are sampled on silicon wafers using electrostatic precipitation with a time resolution of 1–8 h. After collection, the wafers were analyzed on site in an isostatic diffusion chamber. The sample is first cooled to the desired sub-zero temperature, which is measured with an accuracy of ±0.2 ∘C. Then, the sample is exposed to a defined amount of water vapor (i.e., here corresponding to 101 % RHw), relevant for the formation of MPCs. Subsequently, ice crystals start to grow on INPs, which causes a change in brightness that is detected and recorded by a camera.

The INSEKT samples were collected on precleaned (10 % H2O2 solution followed by ultrapure water) polycarbonate filters with pore sizes of 0.2 µm (Whatman plc, Kent, United Kingdom) with a flow of 9.6 Lmin-1 for 8 h (daytime) or 12 h (nighttime). The samples were shipped to the laboratory at the Karlsruhe Institute of Technology for analysis. In brief, the collected aerosol particles are washed off the filter with 8 mL of nanopure water, and the resulting suspension is then pipetted into 80 wells of polymerase chain reaction (PCR) plates, with each well having a volume of 50 µL, together with aliquots from the nanopure water used to create the suspension. The samples are then placed into aluminum holders which are cooled using a cooling bath, measuring down to temperatures of ∼-25 ∘C at a 0.25 ∘Cmin-1 cooling rate. The frozen wells are detected optically by a camera due to a brightness change in the frozen aliquots. For each experiment, a background correction is applied, taking into account the freezing of the pure water used to create the suspension.

The calculation of the cumulative [INP] is based on the sampled volume during filter collection and follows the equation from Vali (1971). Here, we present [INP] at -10, -15, and -20 ∘C for results from INSEKT and -20, -25, and -30 ∘C for FRIDGE ([INP]-10 to [INP]-30).

Aerosol particle composition was measured with three different aerosol mass spectrometers: the laser ablation aerosol particle time-of-flight mass spectrometer (LAAPTOF, AeroMegt GmbH), the Aircraft-based Laser ABlation Aerosol MAss spectrometer (ALABAMA), and a compact time-of-flight aerosol mass spectrometer (C-ToF-AMS).

The LAAPTOF and ALABAMA are based on similar measurement concepts: aerosol particles are transmitted from ambient air into a vacuum chamber through an aerodynamic lens which focuses the particles between 0.07–2.5 µm onto a narrow beam. The particles are then detected by two continuous wave lasers, which allows for measuring the velocity and thereby their vacuum aerodynamic diameter. Particle detection at both detection stages triggers an ablation laser that emits a laser pulse onto the particle, thereby ablating the particle and ionizing its components. The ions are detected in bipolar time-of-flight mass spectrometers. The overall detection efficiency (combining the detection efficiency and the hit rate) of the LAAPTOF is between 0.01 % (±0.01%) and 4.2 % (±2.4%), in the size range of 0.2 to 2 µm based on polystyrene latex (PSL) particles. The highest overall detection efficiency is for 1, and the lowest is for 2 µm (Shen et al., 2018). Note that such efficiency is also particle type dependent (Shen et al., 2018, 2019). More details on the LAAPTOF can be found in Gemayel et al. (2016), Reitz et al. (2016), and Shen et al. (2018, 2019). Details on ALABAMA have been presented in Brands et al. (2011), Roth et al. (2016), Schmidt et al. (2017), and Clemen et al. (2020). The detection efficiency of ALABAMA during this campaign was between 40 % and 60 % in the size range of 0.3 to 1.0 µm based on PSL particles. Up to a particle size of about 1.3 µm, the detection efficiency of ALABAMA decreased to less than 30 % and is estimated to be about 5 % (±5%) for 2 µm. At the same time, the hit rate during those tests using PSL particles was lower, such that the overall detection efficiency for ALABAMA was only between 1 % and 16 % in the size range from 0.3 to 1 µm. As those values are based on measurements using PSL particles, they can vary considerably during field applications; for example, the ALABAMA hit rates were significantly higher than those of the PSL test measurements (which is attributed to particle charge effects during the nebulization of the PSL particles). In light of our research objectives, focusing on the general trend of the aerosol particle composition, we therefore provide an overview of the size-dependent overall detection efficiency from the LAAPTOF and ALABAMA normalized to the maximum value measured, together with the normalized transmission efficiency from HINC (Fig. S1 in the Supplement). From those normalized values it is visible that both SPMSs detect aerosol particles in the same size range, with a maximum between 0.5 and 1 µm, and therefore yield comparable information on the particle composition in this size range. HINC measures particles below 2 µm with a high efficiency, which can have an impact on the comparison between the INP measurements from HINC and the results obtained from ALABAMA and the LAAPTOF.

The data product of both instruments is a bipolar ion signature on a single-particle basis. The main difference between the LAAPTOF and ALABAMA with respect to the chemical analysis is the laser wavelength that is used for ablation. The LAAPTOF uses an ArF excimer laser with a wavelength of 193 nm (pulse duration: 4 to 8 ns), whereas ALABAMA uses a quadrupled Nd:YAG with a wavelength of 266 nm (pulse duration: 6 ns). The laser energy per pulse used during this campaign ranged between 3 and 4 mJ for the LAAPTOF and between 7.2 and 9 mJ for ALABAMA. These differences in laser wavelength and energy result in different ablation and ionization efficiencies that affect the relative ion signals in the particle mass spectra. More specifically, the different lasers lead to different power densities. ALABAMA generates a power density of 1×109Wcm-2, with an assumed effective diameter of the laser at the ablation spot of the particles of about 400 µm (Clemen et al., 2020), whereas the LAAPTOF uses a power density of ∼1×1010Wcm-2, with a laser beam diameter of 99 (±31) µm at the ablation spot (Ramisetty et al., 2018). The C-ToF-AMS has been described in detail in the literature (e.g., Drewnick et al., 2005; Canagaratna et al., 2007). Its data product is a quantitative mass concentration of non-refractory aerosol compounds, e.g., ammonium, nitrate, sulfate, and organics, in particles with an aerodynamic diameter below 1 µm (PM1), based on averaging over an ensemble of particles. The focus of the analysis presented here lies on the single-particle instruments, because they are able to provide information on both non-refractory and refractory particles indicating the presence of mineral dust, elemental carbon, sea spray, metals, or primary biological particles.

For an overview of the ambient aerosol properties of the whole measurement period, we used a clustering algorithm to infer the dominating aerosol particle type. This was done using the LAAPTOF data, which covered a longer time period. ALABAMA was not operated in optimal configuration in the beginning of the campaign because the extraction voltages for the mass spectrometer had to be optimized after the installation of a so-called delayed ion extractor (see Clemen et al., 2020, for more details). Furthermore, during most of the time when the station was in clouds, ALABAMA was connected to the Ice-CVI (see Sect. 2.6) to sample freshly formed ice particle residuals. The mass spectra recorded by the LAAPTOF were classified by fuzzy c-means clustering method as described in Reitz et al. (2016) and Shen et al. (2018). Briefly, this classification method attributes individual particles to multiple classes or clusters according to spectral similarities, typically calculated by linear correlation or Eulerian distance. The particle classes are attributed to particle types based on marker ions that have been identified in reference spectra of particle types.

To show that both the LAAPTOF and the ALABAMA yield comparable results in spite of the different laser wavelengths and energies for ablation and ionization, cation and anion mass spectra from both instruments averaged over 24 h of parallel sampling are compared (see Fig. S2 in the Supplement). Spearman's rank correlation between both mass spectra yields a correlation coefficient of 0.81 for cations and 0.59 for anions. Although many peaks can be identified in both mass spectra, the relative height of many ion signals is different, due to the different ionization laser energy. In general, this suggests a reasonable agreement in the overall detection of aerosol particles between the two instruments.

To evaluate the chemical information from the single-particle mass spectrometry against the [INP]-31 and ns, we did not use the cluster analysis, because, as explained above, cluster algorithms group mass spectra by similarity. This similarity is calculated by linear correlation or Eulerian distance and is therefore mainly influenced by large ion signals. Thus, small ion signals that might still represent a chemical component that is important for ice nucleation may be overlooked by this method. Instead, we chose to correlate the time series of individual ions with the time series of the [INP]-31 and ns as follows: the HINC INP data represented average values over 20 min. To transpose the mass spectrometer data onto these 20 min intervals, we used two methods. Method 1 counts whether a certain ion is present in a mass spectrum during the 20 min interval (i.e., above the noise threshold) and reports the fraction of mass spectra that contain this ion. For example, a fraction of 0.1 means that 10 % of all mass spectra in this 20 min interval contain this ion. Method 2 averages all mass spectra in this 20 min interval after normalizing the spectra to their total ion signal and reports the relative height of a certain ion in the average mass spectrum. Thus, a fraction of 0.1 means that this ion has a relative signal of 10 % with respect to the total ion signal of the 20 min averaged mass spectrum. Method 1 therefore still contains the single-particle information, reflecting the relative abundance of ions on a particle-by-particle basis, whereas method 2 is more representative of the average composition of all particles during this time interval. We note that the sensitivity of the instruments can vary substantially for different ions, which can have a significant impact on the apparent average composition.

The resulting time series of all ions up to m/z 250 were then correlated to the time series of the [INP]-31 and ns using Spearman's rank correlation. This was done for the positive ions (cations) and for the negative ions (anions) separately. Non-significant rank correlations (p>0.05) were excluded. The remaining rank correlation coefficients (r) for all ions up to m/z 250 were squared (r2) and averaged, and their standard deviation (σ) was calculated. Only ions whose r2 values were greater than 1σ above the mean value were selected as meaningful correlators (see Fig. S3 in the Supplement) to [INP]-31 or ns. To correlate the peaks from the particle spectra with [INP]-31 and ns, we used the non-squared coefficients (r) of the selected ions to allow investigation for negative correlations. We note that the instruments used here for the correlation coefficient analysis have different size-dependent particle transmission efficiencies but overlap in the region <2 µm. While HINC, the LAAPTOF, and ALABAMA have their maximum transmission and detection efficiency for particles <1 µm, HINC has a higher transmission efficiency for larger particles (56 % for 2 µm particles), compared to the LAAPTOF and ALABAMA (∼0.01 % and ∼5 %, respectively). The advantage of the ion correlation method is that it looks at the correlation of chemical substances rather than whole particle types, which means that fewer initial assumptions have to be made and a cross-particle type approach can be taken.

In the analysis of both instruments, we find that sea-spray-related ions have elevated correlation coefficients with both [INP]-31 and ns (Tables 1 and S1). The chlorine anions 35Cl- and 37Cl- as well as cations with m/z 46 (Na2+), 81 (Na235Cl+), and 83 (Na237Cl+) in the LAAPTOF (Table 1a) and 35Cl-, 37Cl-, 23 (Na+), 62 (Na2O+), 63 (Na2OH+), 78 (K2+), 12 (Mg2+), 108 (Na2NO3+), 181 (KNa2SO4+), and 165 (Na3SO4+) in ALABAMA show positive correlation coefficients (r) between 0.31 and 0.57 (Table 1b). Whether these correlations indicate that sea spray particles act as INPs or that sea spray particles arrive in air masses together with other particles that act as INPs remains open and cannot be verified by this correlation analysis. Particles of marine origin can be ice-active (e.g., Wilson et al., 2015), which is further discussed below by including data on air mass origin (see Sect. 3.3.7).

Mineral dust (Tables 1 and S1) is certainly a particle type that is expected to act as an INP. In general, we find correlation coefficients between ions indicative of mineral dust with both [INP]-31 and ns in the range of 0.32 to 0.60. For example, the cations m/z 44 (SiO+), 56 (CaO+), 57 (CaOH+), and 75 (CaCl+) as well as the anions m/z 60 (SiO2-), 76 (SiO3-), and 77 (HSiO3-) appear in the LAAPTOF correlation table with correlation coefficients between 0.3 and 0.6. In the ALABAMA data set we find cations such as m/z 7 (Li+), 12 (Mg2+), 23 (Na+), 24 (Mg+), 39 (K+), 41 (K+), 40 (Ca+), 133 (Cs+), and 138 (Ba+) and anions with m/z 43 (AlO+), 76, and 77 (SiO3- and HSiO3-) on the list with correlation coefficients above the threshold. Differences between both instruments can be explained by the laser energy differences used for ablation and ionization and are discussed in more detail below (Sect. 3.3.5). Moreover, the instruments do not cover the same sampling time, which can also be an explanation for this finding.

Another interesting group of ions is the pure carbon ions Cn+ or Cn- (Tables 1 and S1). The number of carbon atoms (n) reaches up to 12 for ALABAMA cations (m/z144=C12+), but for anions in ALABAMA only C1- (m/z 12) is found in the best-correlating ion analysis (Table 1b), with correlation coefficients between 0.34 and 0.62. However, it should be mentioned that C1- (m/z 12) in particular is not a unique feature for elemental carbon, but it is also frequently observed in mass spectra of other particle types. In the LAAPTOF data set, higher carbon atom numbers (up to C6-) are found for the anions (r=0.32–0.47), but no cations were detected (Table 1a). Pure carbon ion Cn+ signals can also be contributed by organics; for example, in the LAAPTOF, SOA can generate very intensive C+. However, our finding shows that the other organic ions, also containing hydrogen or oxygen atoms, are almost absent in the results with positive correlation coefficients. This indicates that the pure carbon ion Cn+ signals are mainly from EC. Nevertheless, a few exceptions exist, which are 59 (C2H3O2-) and 73 (C3H5O2-) as well as 42 (CNO-), 58 (C3H8N+), and 76 (C3H10NO+), the latter indicating biological material or amines, but the pure carbon signals are dominant. The increased correlation of particles showing a typical ionic pattern of elemental carbon (see Fig. S9a, presumably soot or other combustion-related particles) with [INP]-31 and ns might indicate a possible connection to heterogeneous ice nucleation. In addition, [INP]-31 also has a correlation coefficient of 0.5 with eBC, which is slightly higher compared to correlation coefficients for meteorological parameters and aerosol particle concentration <0.1 µm (Fig. S6). We note that this does not necessarily translate to elemental-carbon-containing particles being ice nucleation active, but other species transported in the same air mass could possibly be the INPs (see also Sect. 3.3.7). Laboratory studies of BC particles produced from synthetic fossil fuel are shown to be poor INPs in the immersion freezing mode for temperatures above -38 ∘C (Mahrt et al., 2018; Kanji et al., 2020, Friedman et al., 2011; Chou et al., 2013), which questions the role of pure fossil fuel BC in ice nucleation in MPCs. Recent studies indicate the regional importance of BC particles from biomass burning events (Schill et al., 2020) and are related to mineral phases of the biomass burning particles (Jahn et al., 2020). Interestingly, BC being a tracer with appreciable correlation with [INP] has also been reported for measurements conducted in the winter in the boreal forest at Hyytiäla (Paramonov, 2020). Thus, our findings do not necessarily suggest that EC-containing particles are ice nucleation active, but other internally or externally mixed particles could potentially be the INPs, revealing that there was an anthropogenic impact on ice nucleation during this observation.

Although r2 was chosen to determine which ions are meaningfully correlated with [INP]-31 and ns, we consider r to also identify negative correlations. Here, we find two organic cations in the LAAPTOF data set (m/z 13, CH+ and m/z 29, C2H5+), but several sulfate-related anions in the ALABAMA data: m/z 96 (SO4-), m/z 97 (HSO4-), 99 (H34SO4-), 141 (CHO2SO4- or C2H5OSO4-), 155 (C2H3O2SO4- or C3H7OSO4-), 177 (HSO4SO3-, CH3SO3H2SO3-), 195 (H2SO4HSO4-), 217 (MgH(SO4)2-, Na(HSO4)2- or NH4NaSO4SO3-), and 233 (KSO4H2SO4-). Some of the listed ion identifications are tentative and have been reported in the literature (e.g., Froyd et al., 2010), but all represent possible sulfate-containing fragments being of inorganic or organic sulfate compounds. Such compounds are typically water soluble and are therefore not expected to contribute to immersion freezing INPs (e.g., Pruppacher and Klett, 1997; Cantrell and Heymsfield, 2005; Kanji et al., 2017) at conditions assessed in HINC. Moreover, we do not expect that those components reduced the ice nucleation ability of the ambient aerosols by a coating effect in the immersion freezing regime (e.g., Kulkarni et al., 2014; Kanji et al., 2019).

In most of the cases we find agreement between the two mass spectrometer ions that showed a relationship to [INP]-31 and ns. There are two noteworthy exceptions: (a) mineral-dust-related ions and (b) sulfate-containing ions. There are more mineral-dust-related ions clearly correlated with [INP]-31 and ns in the LAAPTOF data set than in the ALABAMA data set. A likely reason for this is the different particle size and particle-type-dependent detection efficiency of the two instruments (see Sect. 2.3.1). As mentioned before, the LAAPTOF employs a higher power density than ALABAMA, resulting in a higher degree of fragmentation in the mass spectra of the LAAPTOF. These smaller organic fragments interfere less with mineral-dust-related fragments. Thus, interferences between organic fragments and mineral-dust-related fragments, e.g., m/z 40 (Ca+ or C3H4+), 44 (SiO2, C3H8+ or CO2+), and 57 (CaOH+, C4H9+, C2O2H+, C2H5CO+, C3H5O+), are more likely to occur in the ALABAMA mass spectra, as these organic ions are composed of longer chains of atoms and are thus more likely to undergo fragmentation in the ablation and ionization process. The time series of the mineral-dust-related ions detected in ALABAMA can therefore be superimposed by the organic ions. As our data show almost no correlation between organic ions and [INP]-31 or ns, this explains the finding that the apparent correlation between mineral-dust-related ion and [INP]-31 or ns is weaker in the ALABAMA than the LAAPTOF data set. Possible reasons for differences in detecting sulfate-containing ions could be the higher power density of the LAAPTOF, which results in smaller fragments; thus it is weaker in identifying organosulfates, e.g., m/z 141 (CHO2SO4- or C2H5OSO4-) and 155 (C2H3O2SO4- or C3H7OSO4-), and large inorganic sulfate ions, e.g., 195 (H2SO4HSO4-), 217 (MgH(SO4)2-, Na(HSO4)2- or NH4NaSO4SO3-, and 233 (KSO4H2SO4-) than the ALABAMA.

In general, we find that correlations between time series of ions and [INP]-31 are higher than for ns (Table 1, Figs. S7 and S8). This does not imply that ice activation is not controlled by the aerosol particle surface, but it could imply that the particle composition plays an important role. The calculation methods can also contribute to this: ns is derived by normalizing [INP]-31 to the total available aerosol surface area (see Eq. 1). Thus, at high aerosol load it can be that ns will decrease (or stay constant) while [INP]-31 increases only marginally compared to the total aerosol surface, thus potentially leading to a better correlation between the ions and [INP]-31. On the other hand, both methods 1 and 2, which were used to calculate the time series of ions, are normalized methods: method 1 is normalized to the total number of recorded single-particle mass spectra in the time interval (20 min). Method 2 is normalized to the total signal height of the spectra. Thus, both methods are not sensitive to the total aerosol particle number, surface, or mass concentration in the way that is used in the calculation of ns. Further, we find that the highest correlation coefficients between the time series occur using method 2 and [INP]-31. The ion fraction in method 2 represents the average composition of all particles during the 20 min time interval, whereas method 1 reports the fraction of particles that contain a certain ion. The higher correlation coefficient found between [INP]-31 and averaged chemical compositions, compared to the fraction of single-particle ions present in the 20 min time interval reported by method 1, suggests that not only the presence of a substance causes ice activation, but that the relative amount of the substance in the averaged particle ensemble is also important to the ice activation properties of the particle ensemble. The higher correlation with [INP]-31 than with ns supports our initial statement that a non-negligible influence of particle composition impacts ice nucleation. Thus, not only aerosol physical properties such as size or surface area, which are typically dominant predictors of [INP] (e.g., DeMott et al., 2010), but also chemical composition are important.

To illustrate the findings discussed above, we summed the time series of the best correlating ions of a specific particle type to create a proxy for a more general particle type. For this, we chose correlation coefficients with [INP]-31 based on method 2, as this combination yielded in general the highest correlation coefficients. For example, for the sea spray time series we added the time series of m/z +23, -35, -37, +62, +63, +78, and +165 (ALABAMA). Similarly, we also created a time series for sea spray from LAAPTOF data, for mineral dust and for elemental carbon (ALABAMA and the LAAPTOF), and for sulfate-containing particles (ALABAMA only) using the marker ions of the respective particle types listed in Table S1 (bold marked). The signals at m/z +12, +24, +39, and +41 were not considered for the sea spray and mineral dust time series, because these are very common signals across all particle types. The ions at m/z +12 and +24 can be attributed to both carbon and magnesium. Although m/z +12 clearly shows an increased intensity in the EC type compared to the other particle types listed, it is less clear for m/z +24. Therefore, m/z +24 was not used as a marker for any of the particle types. The signals at m/z +39 and +41 are mainly indicative of potassium, which is a common component of mass spectra due to its ionization energy. For example, potassium occur in biomass burning particles (e.g., Silva et al., 1999) but is also present in sea water and in mineral dust. Thus, the ions m/z +39 and +41 were not considered here in the analysis.

These time series are displayed in Figs. 6 and 7, along with the [INP]-31 time series and particle source region information. For the latter, we used emission footprint sensitivities derived from FLEXPART. We summarized the North African and Middle East countries to account for possible source regions of mineral dust and the marine surfaces to account for possible sea spray particle sources (as explained in Sect. 2.6). Figure 5a shows that the time series of the footprint emission sensitivity calculation for dust surfaces for the North African and Middle East countries are elevated during times when [INP]-31 is high and are accompanied by elevated mineral dust marker values, e.g., on 23, 27, and 29 January and 3, 12–17, and 19 February (Fig. 5b). Moreover, the mineral dust proxies from both mass spectrometers agree well (Fig. 5b). The correlation with the [INP]-31 time series has already been discussed above and indicates that mineral dust particles transported from source regions in North Africa and the Middle East to JFJ act as INPs at -31 ∘C.

The time series of the EC proxies from ALABAMA and the LAAPTOF are shown in Fig. 5c. Here, there is some discrepancy between the instruments, for example during the period between 14 and 19 February. Then ALABAMA data show a higher contribution of EC than during the first part of the campaign (29–31 January), whereas in the LAAPTOF data the EC contribution remains similar when comparing 23 January–7 February and 14–19 February, which might be related to differences in the instruments (see Sect. 3.3.5). Thus, the elevated [INP]-31 during 14–19 February occurs during times of not only elevated dust but also elevated EC concentrations. As mentioned before, many studies have reported EC particles do not contribute to heterogeneous ice nucleation above -38 ∘C (e.g., Kupiszewski et al., 2016; Vergara-Temprado et al., 2018b; Mahrt et al., 2018; Kanji et al., 2020). It is possible that the air masses transporting mineral dust from the North African and Middle East source regions also transport polluted air bringing EC-containing particles along the transport pathway. This may result in the observed correlation between INPs and EC-containing particles without the latter being active as INPs and is supported by an increase in the EC signal occurring simultaneously (ALABAMA; see Fig. 5b and c) with an increase in the dust markers on 31 January and 14–18 February.

Figure 6 shows the time series of the relative contribution of marine surfaces from FLEXPART, together with [INP]-31 (Fig. 6a), and the proxies for aged sea spray particles from both ALABAMA and the LAAPTOF (Fig. 6b). Both measurements from the single-particle mass spectrometers show similar trends, although the ALABAMA data show much less variation in the second part of the campaign (after 11 February) than the LAAPTOF data set. The differences in the ALABAMA time series between the first and second halves of the campaign are also partly due to changes in the instrument configurations (as described in Sect. 2.3.2). However, the time series of the sea spray proxies from the mass spectra do not match the FLEXPART emission footprint sensitivities for open ocean and sea surfaces (Fig. 6a). Also, the [INP]-31 time series does not match the open water source regions from FLEXPART. Interestingly, the calculated air mass travel times from FLEXPART between the marine boundary layer and the measurement site reveal similarities with the sea spray proxies. For example, the period when the single-particle mass spectrometers recorded a stronger signal from sea spray particles (2 to 5 February) coincides with comparatively short travel times of the air masses to JFJ (12 to 24 h), mainly originating from the marine boundary layer over the western Mediterranean Sea (Fig. S11 in the Supplement). The short travel time is in line with a reduced impact of sedimentation losses, wet removal, and mixing/dilution with other air masses of different origins. The FLEXPART analysis shows that travel times from the marine boundary layer to JFJ can vary between approximately 24 and more than 96 h. However, we should point out that the FLEXPART model is not able to separate coastal regions from open ocean for this analysis, such that the open water classification used may not be the best way to represent sea spray source regions which tend to occur along shorelines during wave breaking and as a function of wind speed (e.g., Vergara-Temprado et al., 2017).

Apart from this, both the sea spray and the [INP]-31 time series show a similar temporal evolution as the dust source region emission sensitivities (Fig. 5). A possible explanation for this might be that the dry saline lake beds in the deserts have been identified as source regions for salt particles (Prospero et al., 2002). For example, Formenti et al. (2003) observed that airborne particles originating from the Sahara that were sampled off the North African coast were a mixture of aluminum-silicate-based minerals and NaCl-bearing salts. This may well explain our observed correlation between dust particles, NaCl-containing particles, and INPs.

Figure 6c shows the time series of the sulfate proxy ions detected by the ALABAMA. For sulfate ions, we observed a negative correlation with [INP]-31 (see also Table 1). Especially in the first half of the campaign, it can be seen that the sulfate signal in the ALABAMA is high at times when [INP]-31 is low (28–30 January, 2–3 February, 7 February, and 9 February). After 11 February, when [INP]-31 values are higher along with the mineral dust signals and the emission sensitivities of the dust source regions, the sulfate ion time series shows only slight fluctuation and shows no similarities to the [INP]-31 time series. Furthermore, no similarities with the dust source region emission sensitivities are observed. A comparison of the ALABAMA sulfate ion time series with the sulfate fraction measured with the AMS (Fig. 6c, right panel) shows a good agreement in the time points of the variations, although not in the strength of the fluctuations. Thus, the time series of the ALABAMA and AMS sulfate measurements are anticorrelated to the [INP]-31 time series, which is not surprising given soluble aerosol particles are not expected to contribute to heterogeneous ice nucleation above -38 ∘C.

Cloud residuals of 16 separate cloud events were sampled by the Ice-CVI, sub-divided into 43 cloud periods by the shape of the residual particle size distribution. This was done on the basis of the observed concentration ratio R of the size range 300–600 to 80–100 nm: 2R=N300-600N80-100. For R, three types of cloud periods were defined: R<1, 1<R<3, and R>3. R>3 denotes cloud periods where the IPR number size distribution shows a clear minimum or at least a strong decrease at around 100 nm (Fig. S12 in the Supplement), whereas R<1 denotes cloud periods where the relative number of small IPRs is very high compared to larger IPR sizes (Fig. S13 in the Supplement). The cloud period type 3>R>1 is in between and thus rather inconclusive and therefore not included in the further evaluation. Considering sufficient counting statistics based on the UHSAS and the ALABAMA measurements, 20 cloud periods remained for the condition R>3 and R<1 (Tables S3, S4, and S5 in the Supplement), corresponding to 12 separate cloud events.

Since it is very unlikely to find INPs at a diameter below 100 nm (e.g., Vali, 1966; Pruppacher and Klett, 1997; Richardson et al., 2007; Cornwell et al., 2019), this feature of small IPRs is attributed to the presence of secondary ice in the examined clouds. Therefore, the cloud periods presented in Fig. S12 and Table S3 as well as Fig. S13 and Table S4 are interpreted as MPCs with less and more impact of secondary ice production, respectively. This assumption is further supported by the fact that on average the cloud periods interpreted as being more strongly influenced by secondary ice contribution have significantly higher IPR number concentrations. Such an implication of small cloud residual sizes (<200 nm) on secondary ice formation is also reported for measurements conducted at the Zeppelin observatory (Ny Ålesund) in the Arctic (Karlsson et al., 2021). During these periods, the IPR concentration is also higher in the IPR diameter range larger than 300 nm. This implies that residuals from secondary ice particles can also contribute to the size range of IPRs expected from primary ice particles. Figure 7a shows the mean IPR number size distribution for both cloud categories averaged with respect to the duration of the single cloud periods.

The IPR mass spectra recorded with the ALABAMA downstream of the Ice-CVI were analyzed separately for cloud periods with lower and higher contributions of small IPRs as defined in Sect. 3.5.1 (see Tables S3 and S4). During the 20 analyzed MPC periods, a total of 640 IPR mass spectra were recorded with the ALABAMA. These spectra were clustered using a fuzzy c-means algorithm (Roth et al., 2016; Hinz et al., 1999; Bezdek et al., 1984), and the resulting clusters were assigned to particle types based on marker ions and reference particle mass spectra, resulting in eight different particle types. Figure 7a and b show the IPR size distribution for cases of lower (Table S3; R>3) and higher (listed in Table S4; R<1) contributions of small IPRs, respectively. The particle number fractions of those two cloud conditions are depicted in Fig. 7c, and the size-resolved particle number fraction of the two cloud conditions is shown in Fig. 8. The IPR particle types will be discussed in the following. Additionally, a temporal evolution of the particle composition during a selected cloud event is shown in the Supplement (Figs. S14–S16).

Mineral dust and biological particles. In agreement with previous IPR studies conducted at the JFJ station (Kamphus et al., 2010; Worringen et al., 2015; Schmidt et al., 2017; Eriksen Hammer et al., 2018), IPRs of mineral origin were detected in almost every cloud event and were the most frequently detected IPR type by number (58 %, also including Al type particles; see discussion below about “Aluminum and other metal containing particles”), even though ambient temperatures during the campaign were relatively high (>-17 ∘C). Activation of ice by mineral dust at such relatively high temperatures might be caused by biological material on the surface of dust particles, because many IPRs contained ions indicating biological material (see Fig. S17 in the Supplement). Another possible explanation for this relatively high activation temperature could be that the mineral dust particles were highly ice-active dust particles, such as potassium-rich feldspars, which generally nucleate ice more efficiently and at higher temperatures than other mineral dust particles (Harrison et al., 2016). Moreover, the fraction of mineral dust containing IPRs does not appear to be related to the proportion of small IPRs, as it is found in both cloud period classifications. Moreover, mineral dust IPRs were detected in size ranges between 250 and 1500 nm (Fig. 8). Interestingly, the majority of dust IPRs are submicron in size. Particles dominated by ion signals indicating biological particles were also consistently found in both cloud period classifications, but to a general smaller proportion (<10 %), which was also reported by Schmidt et al. (2017) and Kupiszewski et al. (2015).

Aluminum and other metal-containing particles. The aluminum particle type fraction is the second largest particle fraction, contributing more than 25 % by number during cloud periods with a smaller contribution from small IPRs and more than 15 % during periods of high contributions from small IPRs. Although aluminum ions in single-particle mass spectra can indicate mineral dust, Eriksen Hammer et al. (2018) identified a subset of IPRs containing aluminum as possible contamination from the sampling setup, despite all surfaces of the Ice-CVI being coated with a nickel alloy prior to the 2017 campaign. However, in this work the aluminum particle type detected by the ALABAMA contains additional ion signals of Na+, Cl-, cyanide (CN-), and cyanate (CNO-); thus we regard them as real ambient particles and count them as mineral dust particles (see discussion above) rather than sampling artifacts. Such aluminum-dominated particles have also been observed in IPRs by Schmidt et al. (2017) using the ALABAMA and by Kamphus et al. (2010) using an aerosol time-of-flight mass spectrometer (ATOFMS, TSI model 3800) in combination with the Ice-CVI at JFJ in winter 2013 and in winter 2007, respectively. Other metallic particles were detected with a similar fraction (below 10 %) during times of low and high small IPRs, indicating that metal-containing particles act as INPs, which agrees with previous findings (Kamphus et al., 2010; Eriksen Hammer et al., 2018; Schmidt et al., 2017).

Sodium chloride particles. Another frequently observed particle type is dominated by NaCl, indicating sea spray particles (e.g., Sierau et al., 2014). It should be mentioned that the described correlation between sea spray particles and INPs (see Sect. 3.3.1) refers to total sea spray. In the following, we distinguish between aged sea spray (derived from ions such as Na2O+, Na2O(H)+, Na2NO3+, Na3SO4+, and NaSO4-; Gard et al., 1998; Sierau et al., 2014) and pure NaCl mass spectra (containing only Na+, Cl-, and various NaCl compounds such as Na2Cl+, Na3Cl2+, and NaCl2-). The main difference between the IPR composition during the two cloud period classes is the fraction of the pure NaCl particle type, which is clearly higher (about 20 %) during the cloud periods with a high concentration of small IPRs (R<1). This indicates that these pure NaCl particles do not originate from freshly formed small ice crystals as INPs, but rather result from sampling ice crystals formed by secondary ice production. The same effect may explain the previously reported observations of NaCl particles in IPRs (Kamphus et al., 2010; Eriksen Hammer et al., 2018). This interpretation is further discussed in Sect. 3.5.3.

Size-resolved chemical fraction and internal mixtures. The size distribution of the IPRs reveals a dominant submicron fraction (Fig. 8). This challenges the assumption that most of the INPs are supermicron in size (e.g., Mason et al., 2016) and might be related to the sampling location in the free troposphere, where particles typically undergo long-range transport and sedimentation processes prior to being sampled. The size-resolved examination of the particle fractions (see Fig. 8) shows that IPRs containing mineral dust were detected in all size channels between about 150 and 3000 nm (vacuum aerodynamic diameter). IPRs with biological and elemental carbon ion signals are mainly found in the size range between about 150 and 600 nm, whereas the pure NaCl particles have the highest contribution between about 400 and 1000 nm. We note that the attribution of the IPR spectra to particles types is based on the dominant ion signals of the mass spectra and is not an absolute assignment to a pure particle type. Thus, almost all IPR types also show traces of internal mixtures. For example, over 90 % of the IPRs contain signals of sodium (Na+), over 80 % contain signals of potassium (K+), and over 70 % contain signals of Cl- and CN-. Mineral dust particles in turn contain a variety of other signals that indicate sea spray and biological and biomass burning substances. Therefore, the IPR types cannot be considered completely independent of each other and represent internal mixtures.

At JFJ, two main local wind directions prevail due to the topography around the station. To study the influence of wind direction on IPR composition, we selected air masses coming from either the southeast or northwest to the JFJ station and sorted the cloud periods of Table S4 (those with a high contribution of small IPRs, R<1, indicating more impact from secondary ice formation) for these two main flow directions (Fig. 9). Figure 9a and b show the size distribution of the IPR for the two dominant wind directions, and panel c depicts the IPR composition. The fraction of the pure NaCl IPR type is significantly higher (>40 %) for the northwesterly flow than for the southeasterly flow (about 3 %), whereas the other IPR types show only slight differences between the two cases. Pure NaCl particles are generally considered poor INPs and freeze homogeneously or heterogeneously at temperatures below -38 ∘C (Kanji et al., 2017, and references therein) and thus should not be expected to occur in primary IPRs under the temperature conditions at JFJ. If the ice crystals would have originated from homogeneous freezing at higher altitude and settled to JFJ, they would have grown to sizes too large to be sampled with the Ice-CVI, which only collects freshly (small) nucleated ice crystals. We can also rule out that the observation of pure NaCl particles in the IPR population is due to collected cloud droplets based on the working principle of the Ice-CVI, which impacts supercooled cloud droplets on a cold plate where they freeze upon collision. If the Ice-CVI would also transmit liquid cloud droplets, one would also expect an enhanced abundance of sulfate and nitrate signals, which were detected in cloud droplet residuals (Roth et al., 2016; Kamphus et al., 2010). However, this was not the case in our work. Further, scavenging is unlikely, because firstly, the Ice-CVI samples only small, freshly nucleated ice crystals that have been growing only by diffusion (5–20 µm; Mertes et al., 2007), and secondly, scavenging would also apply to sulfate, nitrate, and organic particles that dominate the aerosol population at JFJ (see Fig. 2).

Thus, we conclude that the pure NaCl IPR type is associated with secondary ice processes. At JFJ, the different orographic terrain profiles northwest and southeast of the station lead to stronger vertical wind speeds for the steep rise on the northern side of the research station, resulting in high peak supersaturations (Hammer et al., 2014). This is expected to favor the occurrence of persistent MPCs, compared to the gentler slope on the south of the measuring station (Lohmann et al., 2016). These stronger updrafts at the northwesterly slope can transport larger droplets to higher regions with lower temperatures, which in turn increases the formation of secondary ice crystals (Lohmann et al., 2016; Korolev and Leisner, 2020). Various secondary ice processes, such as ice fragmentation due to surface and spicule bubble bursting during thermal shock, fragmentation during ice–ice collision, or splintering during riming (Hallett–Mossop process), are favored by freezing of large droplets, although the Hallett–Mossop process has its maximum secondary ice production rate at higher temperatures than those occurring most of the time during our measurements (Hallett and Mossop, 1974; Korolev and Leisner, 2020). Furthermore, blowing snow or fluxes of surface hoar frost are also possible sources of secondary ice (Vali et al., 2012; Farrington et al., 2016; Lloyd et al., 2015; Beck et al., 2018). However, no correlation was observed between horizontal wind speeds at JFJ and IPR composition. We note that the sulfate fraction in the IPR composition during times of an increased influence from secondary ice crystal formation (during small IPR samples) is low as for NaCl, although one would expect that sulfates are also present in droplets which contribute to secondary ice crystals. This could be explained by the original cloud condensation nuclei (CCN) mass being distributed over many small ice fragments upon fragmentation of the ice crystal. Thus, the sampled IPR is smaller than the initial CCN. This shifts the IPR size distributions to smaller sizes, such that sulfate-containing IPRs are possibly too small to be detected in the ALABAMA size range, and a preferential sampling of NaCl particles is detected. Indeed, the size distribution from ambient sea spray particles shows a maximum >1 µm (Fig. S10a) while the IPR NaCl particles under northwest wind conditions have a maximum <1 µm (Fig. 10b). Figure S10 also shows that the sulfate-containing aerosol particles in the ambient air have a maximum at smaller sizes than NaCl-containing particles. Moreover, the faster updraft velocities lead to a shorter lifetime of the cloud droplets before freezing, such that aqueous-phase reactions that lead to aged sea spray particle signatures (e.g., Na2NO3+ and Na3SO4+; Gard et al., 1998) are less likely to occur compared to a slow updraft cloud formation situation, which supports the dominance of the small IPRs with pure NaCl signatures. Which process is ultimately responsible for the secondary ice formation with the simultaneous occurrence of pure NaCl IPR cannot be determined from our measurements.