In this part of the series we present ice nucleation measurements of 18 dust samples. Seven airborne samples were collected after advection from the Sahara. Four of the airborne samples were collected directly from the air in August 2013 and 2014 at the Izaña observatory in Tenerife, Spain, using a custom-made large cyclone (Advanced Cyclone Systems, S.A.: flow rate of 200 m3 h-1 and D50=1.3 µm, the diameter at which the collection efficiency is 50 %). The remaining three airborne samples were collected by deposition on solar panels or roofs in April 2014 (Crete and Peloponnese, Greece) and on 10 May 2010 (Aburdees, Egypt). Nine samples were collected from the surface in the following deserts: (i) the Atacama desert in Chile; (ii) a location approximately 70 km from Uluru in Australia; (iii) the Great Basin in Nevada and (iv) the Mojave desert in California, USA; (v) a Wadi in the Negev desert, approximately 5 km from Sde Boker in Israel; (vi) dunes in the Sahara, close to Merzouga in Morocco; (vii) dunes in the Arabian desert in Dubai; (viii) the Etosha pan in Namibia, a dry salt pan; and (ix) the Taklamakan desert in China. A map showing the locations is provided in . Before arriving to the laboratory, samples were stored in various ways: Samples collected from the surface were typically stored for several weeks in PET bottles or other plastic containers. Airborne samples were stored in polypropylene tubes and sealed with paraffin wax tape. In the laboratory, all samples were stored in the dark at room temperature in polypropylene tubes after pre-processing (sieving and milling, see below). While changes in the ice nucleating ability due to water uptake, loss of volatile material, or growth of biological material which may occur during storage cannot be excluded, they are assumed to be minor, because the samples were collected and stored under dry conditions, hardly exposed to air, and kept at a lower temperature than at which they were collected. The Israel sample and the Etosha sample are from the same batch as those studied in . The surface-collected samples were sieved with a cascade of dry sieves (Retsch Vibratory Sieve Shaker AS 200) with 32 µm diameter being the smallest cutoff size. Most samples only contained a few weight percent in this size range. The Australia and Morocco samples were milled using a vibratory disc mill (Retsch, model RS1), as the fraction of particles in the sub-32 µm size range was too low for ice nucleation experiments. Particles in the lowest available size bin (32–64 µm) of the Morocco sample were milled. The Australia sample was first sieved with a coarse, millimeter-range sieve to separate any large material, and the remaining smaller fraction was milled. For the Israel and the Atacama dusts, a sieved and a milled sample were included in the study. The Israel sample was first sieved, and part of a sub-32 µm fraction was milled, while in case of the Atacama sample part of the initial, unsieved batch was milled.

To investigate if the ice nucleation activity of the samples is influenced by biological particles internally or externally mixed with the dust or by organic coatings on the dust particle surface, selected samples were heated to 300 ∘C and stayed at this temperature for 10 h before the ice nucleation experiments. At this temperature proteinaceous material, such as bacterial and fungal INPs, should be denatured , and the majority of organic material, such as glucose, is combusted and evaporated and references therein.

Dust particles were dry dispersed using a rotating brush generator (RBG, Palas, model RBG 1000) with N2 (5.0) as carrier gas into a 2.78 m3 stainless steel aerosol reservoir tank via a cyclone that confined the dust size distribution to below D50=2.5 µm. Total particle concentration was monitored using a condensation particle counter (CPC; TSI model 3772). The tank was filled with particles up to a concentration of 1200 cm-3, which decreased steadily to about 300 cm-3 over approximately 10 h. The tank was cleaned before an experiment by repeatedly evacuating and purging it with N2 until the particle concentration decreased to 30–90 cm-3.

The particle size distribution of all samples was measured with a scanning mobility particle sizer (SMPS; TSI; DMA model 3081, CPC model 3010) for mobility diameters (dm) between 12 and 615 nm and an aerodynamic particle sizer (APS; TSI; model 3321) for aerodynamic diameters (daer) between 0.5 and 20 µm. The mobility and aerodynamic diameters were converted to volume equivalent diameter (dve) by assuming a typical dust particle density of ρ=2.65 g cm-3 and optimizing the shape factor χ to receive the best overlap of two size distributions measured by the SMPS and APS. This yielded χ=1.36, which is in the range of earlier studies . Assuming spherical particles, the area size distribution was calculated and fitted with a bimodal log-normal distribution. The mean particle surface area (Ave,w‾) was calculated from the resulting fit for each sample. Four size distributions and the fit parameters for all samples are provided in . During an experiment Ave,w‾ was reduced by between 6 % to 24 % due to a faster sedimentation of larger particles in the aerosol tank. The Great Basin sample was coarser than the other samples and settled out faster. Therefore two refills were necessary, and Ave,w‾ varied by 64 %.

The quantitative mineralogical composition of the bulk dust samples was investigated with the X-ray diffraction (XRD) Rietveld method using a Bragg–Brentano diffractometer (Bruker AXS D8 Advance with Cu Ka radiation). The qualitative-phase composition was determined with the software DIFFRACplus (Bruker AXS). On the basis of the peak positions and their relative intensities, the mineral phases were identified in comparison to the PDF-2 database (International Centre for Diffraction Data). The quantitative composition was calculated by means of Rietveld analysis of the XRD pattern Rietveld program AutoQuan, GE SEIFERT;. Due to the small amount of the dust sample, it was not possible to do a mineralogical analysis of the identical size fraction as in the tank (<2.5 µm). Instead, the entire size fraction of the airborne and the milled samples, and the sub-32 µm fraction of the sieved samples, was used. The measured mineralogical composition is provided in , and that of the additional Tenerife samples is provided in Table . The Tenerife2014_1 sample was additionally measured by Powder XRD (Panalytical XPert Pro) in Bragg–Brentano geometry, equipped with a copper anode providing Cu Ka radiation. A diffractogram was recorded before and after heating the sample to 300 ∘C for 10 h on a silicon sample carrier.

Thermogravimetric analysis (TGA) of six of the dust samples was conducted by gradually heating the dust samples from 40 to 300 ∘C at 10 K min-1 and continuously recording the mass of the samples in a thermogravimetric analyzer (Model Pyris 1 TGA, PerkinElmer). During the temperature scan, the samples were under a constant nitrogen flow of 20 mL min-1.

The morphology of one sample was investigated using scanning electron microscopy (SEM; FEI Quanta 250 FEG, ThermoFisher Scientific).

Attenuated total reflection infrared (ATR-IR) spectroscopy was carried out on an FTIR (Bruker Vertex 80v) equipped with an ATR cell (Pike GladiATR, diamond ATR crystal). The beam path of the spectrometer and the optical parts of the ATR cell are under a vacuum (1.65 mbar) to minimize the influence of water vapor and CO2. The crystal where the sample is placed sits in a cell that is flushed with nitrogen gas before and through the measurement for the same reason. A liquid-nitrogen-cooled MCT detector is used for spectra acquisition. Spectra of the samples were recorded before and after a heat treatment. For the heat treatment the samples were placed for 10 h into a laboratory oven at 300 ∘C. The spectral window was set between 700 and 4000 cm-1, with a resolution of 1 cm-1. Spectra for the Tenerife2014_1, Etosha, and Australia samples were recorded this way. Raman images were recorded on a confocal Raman spectrometer (WITec alpha300 RSA+) using a 488 nm laser, 50× magnification, grating of 600 lines per millimeter, and a laser power of 4.7 mW. For Raman imaging the dust was impacted on a clean aluminium surface as described in and subsequently mapped with the aid of a Piezo XY stage. Raman mappings were carried out before and after the same heat treatment as that for the ATR-IR measurements. The Etosha and Australia samples were mapped with this method.

Ice nucleation experiments were conducted with the portable ice nucleation chamber PINC;. Aerosol particles are sampled from the tank, dried, and introduced into the chamber, where they are layered between two particle-free sheath air flows. Before an experiment, the two chamber walls are coated with a thin layer of ice. During an experiment a temperature gradient is applied between the walls, leading to diffusion of water vapor and heat. A linear gradient of temperature and the partial pressure of water vapor between the walls leads to supersaturation with respect to ice. At a constant aerosol layer temperature, the RH is raised at a constant rate until supersaturation with respect to water of a few percent (RHw=103 %–105 %) is reached. If ice nucleation occurs under these conditions, ice crystals grow on the INPs and are detected in the lower part of the chamber by an optical particle counter. As the ice nucleation mechanisms cannot be identified visually in the PINC we refer to the deposition mode at RHw<100 % and to condensation freezing at RHw>100 %. We use these relative-humidity-based thermodynamic regimes as an operational definition, which does not exclude the possibility of PCF to occur at RHw<100 % . Condensation mode refers here to the conditions above water saturation where full droplet activation prior to freezing cannot be guaranteed.

In the deterministic concept ice nucleation is assumed to take place at so-called ice nucleation active sites on the particle's surface . The probability of such a site being present on a particle and thus of the particle to nucleate ice at a certain temperature scales with the particle's surface area . To account for this dependency and compare the ice nucleation ability of the different dust samples the ice-active surface site density ns was calculated: ns=-ln⁡(1-AF)Ave,w‾≈INPNtotAve,w‾, with the total particle concentration Ntot and the activated fraction AF given by INP/Ntot. The approximation is only valid for AF<0.1, which is the case in this study. As aerosol particles larger than 1 µm were in the size range of the ice crystals and could not be differentiated based on size in the optical particle counter spectra, the ns was corrected by subtracting the average ns at RHi=100 % to 103 %. At these low RHi values no ice nucleation is expected, thus counts in the size range of ice crystals are assumed to be large dust particles.

The ice nucleation activity of the Australia, Atacama milled, Etosha, Tenerife2014_1, Peloponnese, and Morocco samples was additionally measured after they had been exposed to 300 ∘C for 10 h. For these experiments the tank was not used. Instead, dust was dry dispersed using particle-free air from a sonicated flask via a cyclone with a cutoff of 2.5 µm and a diffusion dryer into a PINC, a CPC, an APS, and a SMPS. The unheated samples were additionally measured with the same setup to allow direct comparison. The ns measured with the tank setup, in comparison to the sonicated flask setup for the unheated samples, varied between a factor of 1/3 to 3, which is a good agreement given the limitations of ns as a comparison parameter . Possible reasons for the differences are the sonicated flask breaking up larger entities, thus leading to a different mineralogy per size bin than the rotating brush generator or uncertainties stemming from the use of a different measurement setup. This shows the limits of using ns for such complex, polydisperse samples, which should ideally remove any size dependency. To reduce setup dependent uncertainty, only results measured with the same measurement setup are compared in the following discussion.

Ice-active surface site density was determined for 18 dust samples, of which four are from the Sahara and were collected after atmospheric transport at the Izaña observatory in Tenerife and three were from after atmospheric transport in the Peloponnese, in Crete, and the Sinai Peninsula in Egypt. Figure  shows RH scans at three temperatures for these Saharan samples together with RH scans measured online at the Izaña observatory during the Cloud Affecting particLes In Mineral dust from the sAhara (CALIMA) campaigns in August 2013 and 2014, when the observatory was located in the Saharan Air Layer. The data and description of the location and campaigns are given in . The online and offline measured ns values agree well. Most of the ns curves of the different Saharan samples span an order of magnitude. In Fig.  a scatter plot of ln⁡(ns) at RHw=102 % against the quartz plus feldspar content of the dust samples is shown, which is discussed in more details below. It reveals that the airborne Saharan samples are similar in ns and their quartz plus feldspar fraction (20 wt %–30 wt %). This holds true for most other mineral components, as shown in Table . The biggest differences are found for smectite and calcite, the latter accounting for 25 wt %–33 wt % in the Crete, Egypt, and Peloponnese samples but only for 4 wt %–7 wt % in the Tenerife samples. Smectite, in contrast, was only found in the Tenerife samples (23 wt %–32 wt %) and not in dust from the other three locations. Smectite and calcite are both known to have low ice nucleation activity . Thus, the similar amount of one or the other mineral in all Saharan dust samples is in line with the observed similar ns. Differentiating between microcline and orthoclase and comparing the bulk mineralogy to the surface-dependent ice nucleation ability introduces uncertainty. Within this uncertainty, there is no detectable effect from the presence of microcline versus orthoclase in the airborne samples at the studied temperatures. This is in line with the findings by that there is no correlation between ice nucleating ability and the level of ordering in the aluminosilicate framework, which determines if orthoclase (less ordered) or microcline (more ordered) is present.

To determine how well the dust mineralogy can overall predict the ice nucleation activity, the correlation of ln⁡(ns) with the fractions of the most common minerals in the dust samples was compared in the partner paper . This comparison showed that the immersion-mode ice nucleation activity correlates best with the K-feldspar fraction alone at T=253 K, a temperature where only feldspar minerals are found to have significant ice nucleation activity . At T≤245 K, the ice nucleation activity correlates best with the quartz and quartz + feldspar fractions of the dust samples . At these lower temperatures, quartz as well as, to a lower degree, clay minerals were found to nucleate ice efficiently . Before investigating the role of other compounds, here we do a similar correlation analysis of the ln⁡(ns) at T=238, 240, and 242 K and RHw=97 % and 102 % with the fraction of various minerals contained in the dust samples. Figure  exemplarily shows a scatter plot of ln⁡(ns) at RHw=102 % against the quartz plus feldspar content of the dust samples. The Etosha sample was excluded from the correlation, as it does not contain any significant amount of these minerals. The surface-collected samples with a high ratio of quartz and feldspar tend to have a higher ns than the airborne samples. This trend is similar at the three investigated temperatures. The resulting correlation coefficients for the investigated mineral fractions are provided in Table . The findings by for the immersion mode ns are confirmed in the deposition and condensation mode: at all three tested temperatures, the highest correlation of ns is found for the fraction of quartz, followed by the quartz + feldspar fraction. The correlation with quartz alone is dominated by the Australia sample, which consists of 91 wt % quartz and is by far the most ice nucleation active dust sample. The remaining samples, which consist of, at most, 64 wt % of quartz, correlate only weakly with quartz alone (e.g., R=0.29, p=0.27, at T=238 K and RHw=102 %) and better with quartz plus feldspar (R=0.41, p=0.08). Adding illite to the quartz + feldspar fraction leads to insignificant changes of the correlation coefficient while adding kaolinite reduces it. Calcite and kaolinite alone are always negatively correlated with ns in both deposition and condensation mode. These observations are reasonable when compared to the findings of , who found a higher ns for quartz over clay minerals at T>236 K but lower ns for quartz than for feldspars for T>242 K. Generally, the correlations are lower in this study compared to the immersion-mode data from . A possible reason for this is that the only partial activation of INPs at these measurement conditions in the PINC due to an inhomogeneous RH profile along the particle trajectories inside the chamber weakens the effect of differences in mineralogy. Another reason could be that surface coatings play a more prominent role at lower RH values, because they are less diluted than in immersion mode. Correlating mineralogy, which is based on the bulk fraction, with the surface property ns, leads to additional uncertainty, as described above. Effects by non-mineral substances such as coatings are discussed in the following.

In this section, the role of heat labile material on the surface of dust particles is investigated. A representative subset of the samples was selected to reduce the number of experiments necessary. The Australia and Morocco sample were selected because of their exceptional high ns, the Etosha sample because its mineralogy did not explain the observed ns, the Atacama milled sample because we expected a higher ns from the mineralogy, and finally the Tenerife2014_1 and Peloponnese samples as representatives for two airborne samples from different locations. Figure  shows the ns at 240 and 242 K of the unheated and the heated samples. In case of the Tenerife2014_1 sample the heat treatment led to an increase of ns at RHw<100 %, while it had no effect above water saturation. In contrast, the maximum ns of the Etosha sample decreased by 1 order of magnitude at 240 K after heating. At 242 K the ns was below the detection limit, as can be seen by a scattered flat ns curve in Fig. b. The heating had a small decreasing effect on the ns of the Peloponnese sample at all RH values and little to no significant effect on the ns of the remaining samples. A change in ice nucleation ability due to the heating gives an indication of the nature of the active sites, i.e., if they are inherent to the minerals themselves or a semi-volatile coating material or are due to biological particles, which are impacted by the heating. An RH dependency of the change in ice nucleation ability may suggest that the material which contains or coats the active sites is susceptible to dissolution.

We investigated these possible implications further by thermogravimetric analysis, ATR-IR, and Raman spectroscopy of the samples. The relative mass loss under increasing temperature is shown in Fig. a and its first derivative in Fig. b. If pure samples of single minerals or organic species were studied, the TGA would show discrete steps in mass loss, indicated by spikes in the derivative plot. However, the complexity of the dust samples in this study, which consist of several minerals and likely also various other components, cause the spikes to widen, thus reducing the possibility to observe discrete steps. For example, showed that the mixture of a montmorillonite with varying concentrations of organic surfactant shifts the peak in the derivative mass between the higher temperature of the montmorillonite and the lower temperature of the surfactant, depending on the concentration of the surfactant. This makes it impossible to identify exactly which species are evaporated at which temperature. However, taking into account the temperature range, a qualitative analysis is possible.

The Morocco, Australia, and Atacama milled samples, which all showed no change in ns after heating, exhibit a small, gradual decrease in mass of, at maximum, 0.5 % between 40 and 300 ∘C. In the case of the three samples whose ns changed after heating, i.e., the Etosha, Peloponnese, and Tenerife2014_1 samples, a larger mass loss is found. A first decrease in mass of about 1 % (Etosha) to 2 % (both airborne samples) is observed between 40 and 110 ∘C. The Etosha sample shows a second graduate mass release from 220 ∘C onwards (1 %). The two airborne samples, Peloponnese and Tenerife2014_1, show small steps in mass release between 110 and 225 ∘C and 110 and 170 ∘C, respectively, and a continuous reduction in mass above these temperatures, reaching a total loss of 3 % and 8 % at 300 ∘C, respectively.

The temperature ranges where the mass loss occurs can be related to different materials which were evaporated and potentially altered the ice nucleation behavior. The first decrease in mass at 40 to 110 ∘C is mostly due to the evaporation of adsorbed water on the surface of the dust particles or of volatile material such as volatile organics. At temperatures between 110 and 300 ∘C the mass loss is mainly related to decomposition of organic matter, e.g., amides, carboxylic, and phenolic functional groups or the combustion of certain organic compounds such as cellulose . studied various biogenic organic materials, such as cellulose, glucose, bacteria (Escherichia coli and bacillus subtilis), and humic and glutamic acid, which showed either a bimodal or trimodal pattern in the TGA derivative. For all samples, they observed a first peak at 160–180 ∘C, which accounted for 10 %–42 % of the mass loss and a second mass loss peak between 340 and 490 ∘C. Most minerals present in the dust samples are stable at T≤300 ∘C, apart from smectite and gypsum , as further discussed below.

We performed ATR-IR spectroscopy on the Etosha and Tenerife2014_1 to investigate the nature of the material responsible for the respective decrease and increase in ns with heating. Furthermore, we chose the Australia sample as representative for most other cases where the ns stayed the same. Figure  shows the ATR-IR spectra of the three samples before and after heating. The bands between wavenumber 700 to 1200 cm-1 are related to the dominant minerals in the (bulk) samples: quartz in case of the Australia sample; dolomite, calcite, and ankerite in the case of Etosha; and kaolinite and smectite in case of the Tenerife2014_1 sample. Kaolinite also has signals at 3619, 3650, and 3685 cm-1 . No significant differences are observed between the unheated and the heated Australia and Etosha samples. The intensity differences in the mineral bands are likely related to single grains not representatively amplifying the signal during sampling. The Tenerife2014_1 sample, in contrast, shows clear differences before and after heating. A loss of intensity in the OH stretch region between 3500 and 3100 cm-1 is observed, which is too pronounced to be from adsorbed water only. Further intensity loss is observed in the C–H aliphatic region, with decreasing bands between 3000 and 2850 cm-1. This points to volatile organics being present on the unheated sample which were released during heating.

Raman mapping was performed on the Etosha, Tenerife2014_1, and Australia samples. Due to strong fluorescence, however, the Tenerife2014_1 spectra did not yield any information and are thus not presented here. The Raman maps for the Etosha and Australia samples are shown in Fig. . They reveal bands between 1200 and 1700 cm-1 with a distinctive pattern related to soot and carbonaceous material being present on many particles in the Australia sample (Fig. a2 and a4) and on some in the Etosha sample (Fig. c2 and c4). The carbonaceous material is not affected by the heating (Fig. b and d). A cluster with a strong broad signal at 3180 cm-1 and a secondary band at around 1080 cm-1 is observed for the unheated Etosha sample, which is present on most particles (Fig. c1 and c3). After the heat treatment, no cluster is identified anymore containing the broad 3180 cm-1 band, yet the signal at 1080 cm-1 remains. This supports the interpretation that the band at 1080 cm-1 can be attributed to calcite. Furthermore, the band at 1080 cm-1 does not correlate with the 3180 cm-1 band, therefore they belong to two different compounds. The cluster analysis groups these signals for the unheated sample due to the spatial proximity of the materials on the dust. It is likely that on mineral dust grains with a high calcite content a compound was absorbed. During the heat treatment this absorbed compound disappeared and the calcite remained. Afterwards, the cluster analysis showed no group containing the 3180 cm-1 band. A filter (3180±175 cm-1) was applied to search specifically for this signal. This revealed only very few particles carrying the related material after heating (Fig. d1 and d3).

Identification of the material which was released or decomposed during the heating was hampered by fluorescence inherent to the minerals in the samples and also possibly due to biological material if present in the unheated samples. The ratio of signal to noise (fluorescence) was optimized by impacting small amounts of the samples on a pure aluminum surface and by adjusting the laser power, but the fluorescence could not be entirely suppressed. This, together with the complexity of the samples, inhibited an unambiguous identification of the species which were altered by the heating and may affect the ice nucleation ability. We suggest three possible candidates for the cluster with a strong Raman signal at 3180 cm-1 in the Etosha sample. (a) Amides typically show a Raman signal between 3300 and 3100 cm-1 , as observed in the Etosha sample, but distinct peaks are usually also observed for amides between 1700 and 1600 cm-1, which are absent in the Etosha sample. (b) Pure ammonium sulfate has a broad band above 3000 cm-1 and a sharp peak at around 990 cm-1 . While the broad band at 3180 cm-1 agrees well with the observed signal, a second band is found at 1080 cm-1 in the Etosha sample. Ammonium sulfate has a band at 990 cm-1, which is significantly different from 1080 cm-1. Thus, the signal can be attributed to the calcite which has a very prominent band at this position. In addition, the XRD measurements confirmed calcite to be present in the sample. (c) Cellulose shows a broad band between 3575 and 3125 cm-1 and numerous bands between 1320 and 1030 cm-1 . Additional bands at 1750 and 1725 cm-1, 1635 and 1600 cm-1, and 1480 and 1435 cm-1 are not distinguishable in the Etosha Raman spectrum.

For the Etosha sample an effect of organic or other heat labile material on the ice nucleation behavior appears likely. The main minerals contained in the Etosha sample (i.e., ankerite, calcite, dolomite, and muscovite) are not known to be particularly ice nucleation active at the investigated temperatures. In case of ankerite the ice nucleation ability is unknown. Based on its similarity with dolomite, a carbonate known not to be ice nucleation active, it is assumed that ankerite is also not active. Thus, one of the suggested candidates with the strong Raman signal at 3180 cm-1 is likely responsible for the ice nucleation activity of the Etosha sample. Being part of proteins, amides are ubiquitous in nature. Similarly, cellulose is the most abundant organic compound on Earth amongst others as a structural component of algae. Both cellulose and some proteins have been identified to cause ice nucleation at the studied temperatures. Ammonium sulfate has been observed to correlate with higher INP concentration in Saharan dust and to increase the freezing onset temperature of microcline and various other minerals by up to 3 K. According to a study on the Ntwetwe Pan in Botswana by , the organic carbon concentration in a salt pan is about 1 wt % at the surface and consists of cyanobacteria and algae. We assume that similar values apply also for the Etosha pan. The Etosha sample was collected from the edge of the salt pan, a few hundred meters away from a fertile soil area containing the highest organic carbon content of the national park . As wind erosion was identified in these nearby fertile soils, aeolian transport potentially led to higher organic matter concentration at the edge of the pan compared to farther towards the center of the pan. Overall, this suggests that the Etosha sample's ice nucleation ability is almost entirely caused by organic matter, potentially cellulose or proteins which were mixed with or adsorbed onto the mineral dust on the ground, explaining the almost complete suppression of ice nucleation of the heated samples.

In contrast to the Etosha sample, the Tenerife2014_1 sample consists of a number of minerals ice nucleation active at the studied temperatures, e.g., orthoclase, plagioclase, and quartz (Table ). The sample shows the largest mass loss in the TGA analysis. The rather steep step in the TGA loss curve at about 120 ∘C suggests a certain species to be released at this temperature, probably containing aliphatic compounds as suggested by the ATR-IR measurements. The complexity of the airborne Saharan samples is indicated by , who studied organic material in the Saharan Air Layer. They collected 42 PM2.5 and PM10 filters at the Izaña observatory in parallel to the Tenerife2013 sample collection and the CALIMA 2013 campaign in August 2013. Organic matter accounted for about 1.5 wt % of the aerosol composition in the Saharan Air Layer and was determined to mainly consist of saccharides, related to organic material in soils, biogenic secondary organic aerosol particles resulting from isoprene and α-pinene oxidation, and organic compounds from natural and anthropogenic sources such as vegetation and engine emissions. During daytime, the boundary layer reaches the altitude of the Izaña observatory and organic matter can originate from local sources. During nighttime the observatory is located in the free troposphere and aerosol sources are distant. Anthropogenic and natural emissions can originate from the North African coast or be advected from Europe . As the dust collection took place over several days and nights, an influence of organic matter from local sources cannot be excluded.

Another explanation for the reduction in mass at temperatures below 300 ∘C is the release of free water molecules from the crystal lattice as indicated also in the ATR-IR spectra between 3600 and 3100 cm-1. In some cases, this affects the crystal lattice: Smectite, a swelling mineral, collapses under decreasing water vapor pressure as experienced during the heating. This decreases the interlayer spacing (for an overview of the effect of layer charge on smectite swelling see ). It is unknown if the change in crystal lattice has an effect on the ice nucleation ability of the otherwise only weakly ice nucleation active smectite . An effect cannot be excluded as the lattice match with ice is believed to be one of the factors promoting ice nucleation . In the studied samples, smectite was present in the Tenerife2014_1 sample (23 wt %), and traces were found in the Etosha sample (1 wt %). Thus, the collapse of the smectite lattice should only influence the Tenerife2014_1 sample. In case it has an influence, this would be related to an increase in ice nucleation ability.

XRD analysis of the unheated and heated Tenerife2014_1 sample show the conversion of gypsum to anhydrite (Fig. ). Gypsum has a low ice nucleation ability, similar to the clay minerals kaolinite and illite . found anhydrite to have a higher ice nucleation activity than quartz in the immersion mode at temperatures below 243 K when dry generated but a much lower ns when particles were generated from an aqueous solution. Anhydrite transforms back to gypsum when exposed to a relative humidity higher than 97 % at room temperature . However, this process occurs on the order of hours to days, in line with the observed differences in ns between wet and dry generated particles in . Potentially, the transformation to anhydrite during heating explains the higher ns of the heated Tenerife2014_1 sample compared to the unheated one at subsaturated conditions. In this case, a partial conversion of anhydrite back to gypsum during RH conditions above water saturation might explain the unchanged ns of the unheated and heated Tenerife2014_1 sample above water saturation. It should be kept in mind that the bulk mineralogy as determined by XRD is not necessarily representative of the particle surface where ice nucleation takes place. As needle formation has been observed in the transformation of gypsum to anhydrite , we use the occurrence of needles in our sample as an indication that the gypsum–anhydrite transformation took place on the surface of particles and thus might be responsible for the change in ice nucleation behavior of the sample. In SEM images of the unheated Tenerife2014_1 sample (Fig. a), hardly any needles are visible. A small number of needles is observed at the center of the image of the heated sample (Fig. b), while no needles are found in other SEM images of the heated sample (see the Supplement). The apparently limited needle formation and the fact that only about 1 wt % gypsum is contained in the sample suggests that gypsum transformation under heat treatment should only have a small effect on the ice nucleation behavior of the Tenerife2014_1 sample. However, given that, at maximum, only about 10 % of the particles act as INPs in case of the Tenerife2014_1 sample, the gypsum–anhydrite transformation might be non-negligible. Additionally, we suggest that the increase in the ns found under subsaturated RHw conditions for the Tenerife2014_1 sample is caused by the volatilization of aliphatic compounds containing matter, as indicated by the ATR-IR and TGA measurements, which inhibited the active sites of the mineral dust itself.