Freezing of droplets consisting of liquid water or aqueous solutions and subsequent ice crystal growth affects microphysical and radiative properties of clouds and precipitation. Understanding and predicting the formation of ice in clouds is critical to quantifying the impact of aerosols on climate (DeMott et al., 2011). Ice crystals in the atmosphere may form by homogeneous ice nucleation of cloud droplets at temperatures T<237 K or by heterogeneous ice nucleation under warmer conditions and possibly in the whole temperature range T=237–273 K. For heterogeneous nucleation, ice-nucleating particles (INPs), i.e. particles that initiate a phase transition in an environment supersaturated or supercooled with respect to ice, need to come in contact with (contact freezing), be immersed in (immersion freezing), and act as condensation nuclei for water droplets (condensation freezing), or present a surface for deposition of water vapour from the gas phase (deposition nucleation) (Pruppacher and Klett, 1997). Field measurements have shown that heterogeneous ice nucleation is indeed an important process in the atmosphere since ice formation in cumulus and stratiform clouds begins at temperatures much warmer than those associated with homogeneous ice nucleation in pure water (Korolev et al., 2003; Ansmann et al., 2009; Seifert al., 2010; Kanitz et al., 2011). Recently, Cziczo et al. (2013) suggested that heterogeneous ice nucleation might even have a dominating role in the formation of high, cold cirrus clouds.

INPs are generally rare and may represent less than 1 in 106 of the aerosol particle population (DeMott et al., 2011). Various insoluble particles such as mineral dust, soot, metallic particles, volcanic ash, or primary biological particles have been suggested as INPs (Hoyle et al., 2011; Hoose and Möhler, 2012; Marcolli, 2014). The ability of mineral dusts to nucleate ice has been demonstrated in many laboratory experiments (Hoose and Möhler, 2012 and references therein; Marcolli, 2014 and references cited therein), and their impact on cloud properties has been shown in observational and modelling studies (Lohmann and Diehl, 2006; Hoose et al., 2010; Choi et al., 2010; Ansmann et al., 2008, 2009; Seifert et al., 2010). In mixed-phase clouds, mineral dust particles are usually the dominant INPs (Pratt et al., 2009; Kamphus et al., 2010) and were also found to be important for the heterogeneous freezing of cirrus clouds (DeMott et al., 2003; Twohy and Poellot, 2005; Cziczo et al., 2013).

The main sources of mineral dusts in the atmosphere are the arid regions in the world (Prospero et al., 2002; Ginoux et al., 2012). Model estimates indicate that global mineral dust emissions by wind-driven erosion range between 1000 and 3000 Tg yr-1 (Zender et al., 2004; Cakmur et al., 2006; Textor et al., 2007). The main source is the global dust belt, which stretches from the Sahara to the Taklimakan in China. Sources outside the global dust belt are deserts located in the USA and Mexico, Australia, Botswana and Namibia, Bolivia, and western Argentina (Sassen et al., 2003; Engelstaedter et al., 2006; Formenti et al., 2011). From these source regions the dust may be transported in a wide altitude range over large distances (Wiacek and Peter, 2009; Wiacek et al., 2010; Engelstaedter et al., 2006; Li et al., 2008; Reid et al., 2003).

Immersion freezing is often considered the most important nucleation mode for mineral dusts (Hoose et al., 2010; Murray et al., 2012; Wiacek et al., 2010). The ability of mineral dusts to nucleate ice depends on their mineralogical composition (Murray et al., 2012; Atkinson et al., 2013), but may also be influenced by the presence of coatings or biological material (Möhler et al., 2008; Cziczo et al., 2009; Sullivan et al., 2010a; Conen et al., 2011; O'Sullivan et al., 2014; Tobo et al., 2014; Augustin-Bauditz et al., 2016). The surface structure and therefore the nucleation ability of mineral dust can change due to interaction with organic or inorganic substances. The effect of a coating depends on many different factors such as the mode of freezing, freezing temperature, and thickness and chemical composition of the coating (Cziczo et al., 2009; Chernoff and Bertram, 2010; Sullivan et al., 2010a, b; Tobo et al., 2012). Clay minerals and Arizona test dust (ATD) show a decrease of the ice nucleation ability due to coatings (Cziczo et al., 2009; Chernoff and Bertram, 2010; Sullivan et al., 2010a). Coatings that are acquired by dust particles during transport might be of minor importance for ice nucleation once a liquid cloud has formed and freezing occurs in immersion mode. In the absence of surface chemical reactions, the coating may dissolve in the cloud droplet so that the bare surface is exposed again and the ice-nucleating ability is restored (Sullivan et al., 2010b; Tobo et al., 2012; Kulkarni et al., 2014; Wex et al., 2014). Internal mixing of mineral dust particles with biological material is supposed to increase the ice nucleation ability of the dust and might occur during cotransport, during precipitation, or by lifting of soil particles (Baker et al., 2005; Pratt et al., 2009; Conen et al., 2011; Hallar et al., 2011; Creamean et al., 2013; O'Sullivan et al., 2014; Tobo et al., 2014; Augustin-Bauditz et al., 2016; Wang et al., 2016).

The most abundant components of mineral dusts are typically quartz, clay minerals, feldspars, and calcite (Murray et al., 2012; Boose et al., 2016). Clay minerals such as kaolinite, montmorillonite, or illite have usually been considered the main responsible components in mineral dusts for ice nucleation. Early studies indeed found kaolinite and montmorillonite among other minerals, as inclusions of atmospheric ice crystals (Kumai, 1961; Kumai and Francis, 1962). Many studies used therefore clay minerals as surrogates for mineral dusts (Hoose and Möhler, 2012; Welti et al., 2014; Wex et al., 2014; Wheeler et al., 2015; Augustin-Bauditz et al., 2014; Marcolli, 2014; Hiranuma et al., 2015; and citations therein). Recently, Atkinson et al. (2013) suggested that feldspar particles may account for a large proportion of INPs that contribute to freezing just below about 258 K.

To investigate the role of mineral dust particles as INPs, and the dependence of ice nucleation activity on mineralogical composition, natural dust samples from global dust source regions must be analysed. Dust collected directly from the air as well as dust collected from the ground in regions where dust events occur frequently should be compared with mineral reference samples in order to give new insights into the influence of mineral dusts on freezing in mixed-phase clouds. This study presents freezing results of natural dust samples collected from the ground at reported source regions of atmospheric mineral dust around the globe. To preserve the morphology, the only processing applied to the samples was sieving. We investigated whether the ice nucleation efficiency of the samples show significant differences between different source regions and whether the freezing behaviour can be related to the mineralogical composition. To this end, a set of reference minerals was subjected to freezing experiments using the same procedure as for the natural dust samples.

For the immersion freezing measurements, we used a differential scanning calorimeter (DSC) Q10 from TA instruments. We performed both emulsion freezing and bulk freezing experiments: the former to characterize the average ice nucleation efficiency of dust particles; the latter to determine the temperature at which the best available ice-nucleating particles/sites in a dust sample are active.

For bulk freezing experiments the investigated dusts and minerals were mixed with water molecular biology reagent from Sigma-Aldrich, which proved to have lower average freezing temperatures of droplets compared with our Milli-Q water. Of this suspension, a 1.8–2 mg droplet was placed in an aluminium pan, covered with mineral oil to avoid evaporation or condensation and finally hermetically sealed. The sample was subjected to repeated freezing cycles with 10 K min-1, which is a suitable cooling rate as the freezing temperature is given by a clear heat release onset due to the sudden freezing of the whole droplet. Bulk measurements were performed with a 5 wt % suspension. Blank bulk freezing experiments with the pure water showed freezing usually at approximately 250 K. The highest freezing temperature of pure water was observed at 252.5 K. Freezing at higher temperatures is therefore attributed to the presence of the dusts and minerals in the samples.

For emulsion freezing experiments, a quantity of sieved dust, or a powder of the mineral in the case of the reference samples, was mixed with water molecular biology reagent. Then, 20 vol % of this suspension plus 80 vol % of a mixture of 95 wt % mineral oil from Aldrich Chemical and 5 wt % lanolin from Fluka Chemical were emulsified with a rotor–stator homogenizer (Polytron PT 1300D with a PT-DA 1307/2EC dispersing aggregate) during 40 s at 7000 rpm. Next, 4–15 mg of this emulsion was placed in an aluminium pan, hermetically sealed, and subjected to three freezing cycles following the method developed by Marcolli et al. (2007). The first and the third freezing cycles were executed at a cooling rate of 10 K min-1 to control the stability of the sample. The second freezing cycle was executed at a 1 K min-1 cooling rate and was used for evaluation. The evaluation was done using the implemented software (TA Universal Analysis) of the instrument. The heterogeneous freezing peak was analysed in terms of three characteristic temperatures that correspond to 10 % (Thet,10%), 25 % (Thet,25%), and 50 % (Thet,50%) of heterogeneously frozen water volume. This value was obtained by integrating over the heat flow signal of the heterogeneous freezing peak and setting the total heterogeneous heat flow to 100 %. If the homogeneous and heterogeneous peaks overlapped, the heterogeneous peak was assumed to end where the heterogeneous peak shape was clearly influenced by the homogeneous peak (see Appendix A for further information). The integrals over the heterogeneous and homogeneous peaks are used to calculate the ratio between homogeneous and heterogeneous freezing. Emulsion measurements were performed with 0.5, 1, 2, 5, and 10 wt % mineral dust suspensions. For calcite, dolomite, dolostone, and ankerite it was hardly possible to evaluate the 10, 25, and 50 % heterogeneously frozen water volume because of a very weak heterogeneous freezing signal. Due to this, these minerals were also measured with 50 wt % suspension concentration, and onset temperatures were evaluated. The evaluation method for onset temperatures is described in Zobrist et al. (2006). Very abrupt but small spikes are excluded from the evaluation, because they originate from single droplets in the tail of the size distribution, which are orders of magnitude larger in volume than the average droplets. For dust samples with many large particles (from micrometres upwards), spikes were more frequent. Comparison of the evaluated temperatures for different samples of the same dust with the same concentration usually showed standard deviations of 0.5 K. Samples with weak heterogeneous signal or with spikes overlapping the smooth heterogeneous signal developed standard deviations up to 1 K.

Pictures for the evaluation of the size distribution of emulsion droplets were taken with an optical microscope (Olympus BX-40). Ten different emulsions containing a total of 1700 droplets were analysed. The droplet diameter was evaluated with the free image processing and analysis program ImageTool (from the University of Texas Health Science Center at San Antonio), grouped into bins with a width of 0.5 µm from 0 to 10 µm, and fitted with a lognormal number distribution N(d)=N0ew2/22πwde-ln⁡2(d/dm)2w2, giving dm= (2.41 ± 0.04) µm as the droplet mode diameter, and w= 0.507 ± 0.014 as the mode width, where N0 is the total number of droplets. The volume distribution was fitted with the function V(d)=V0ew2/22πwde-ln⁡2(d/dm)2w2, where dm= (6.1 ± 0.4) µm, w=0.53 ± 0.05, and V0 is the total volume of all droplets. The obtained number and volume distributions are shown in Fig. 1.

The natural dust samples were sieved with a 32 µm sieve. Sieving was performed with the dry samples. No other pretreatment was applied, except for the Antarctica sample for which a milled sample was used in addition to the sieved sample.

Size distributions of the sieved natural dusts and the milled reference minerals were measured with a TSI 3080 scanning mobility particle sizer (SMPS) and with a TSI 3321 aerodynamic particle sizer (APS) and combined as described by Beddows et al. (2010). The bulk powders were aerosolized in a fluidized bed aerosol generator (TSI Model 3400A) to determine the size distributions with SMPS and APS. Similar to the size distributions of the emulsion droplets, also the dust number size distributions were fitted to a lognormal distribution. Results are given in Table 1. Reference minerals were provided by the Institute of Geochemistry and Petrology of ETH Zurich and milled with a tungsten carbide ball mill. No additional treatment was applied.

The mineralogical composition of the sieved natural dust samples and the milled reference samples was measured by X-ray diffraction (XRD). A quantitative analysis was performed with the AutoQuan program which is a commercial product of GE inspection technologies and which makes a Rietveld refinement (Rietveld, 1967, 1969).

Knowing the droplet size distribution of the emulsions and the size distribution of the dust particles, the theoretical value for the latent heat release for homogeneous and heterogeneous freezing can be estimated. This theoretical value can be directly compared with the measured latent heat release by the DSC.

To calculate the total number of mineral dust particles in an emulsion, the average volume of particles V‾p was calculated by V‾p=∫0∞N(r)43πr3dr, where N(r) is the normalized particle size distribution of the respective mineral/dust as a function of its radius r. With the mass of the dust, mdust, and assuming a density of dust, ρdust=2.6 g cm-3 (Möhler et al., 2006), the number of dust particles n in an emulsion can be estimated to be n=mdustsampleρdustV‾p. The probability Pj for a particle to be in a droplet j with a volume Vj is Pj=Vj/Vtot, where Vtot is the total volume of all droplets in the emulsion. Assuming n particles in the emulsion, which are all distributed among the water droplets, the probability for no particle in a droplet j with a volume Vj is (1-Vj/Vtot)n. The contribution of droplet j to the total heterogeneous and homogeneous peak area Atot is proportional to Vj/Vtot. The percentage of homogeneous freezing, phom, can then be written as phom=∑j=1kVjVtot⋅1-VjVtotn, where k is the number of droplets. The fraction of heterogeneously frozen volume phet is then phet=1-phom. Like this, we are able to calculate the heterogeneously frozen water volume under the assumption that all mineral dust particles are able to induce heterogeneous freezing.

The fraction of heterogeneously frozen volume measured with the DSC (phet,lab) is calculated by dividing the latent heat release of the heterogeneous freezing signal by the latent heat release of the total freezing signal (homogeneous plus heterogeneous). The number of mineral dust particles, nlab, necessary to explain the heterogeneously frozen volume fraction can be obtained by solving the equation phet,lab=1-∑j=1kVjVtot⋅1-VjVtotnlab for nlab. The active fraction fact of dust or mineral particles, i.e. the fraction of particles that are active as INPs, can then be calculated by fact=nlabn. Correspondingly, 1-fact is the fraction of dust (or mineral) particles that remain inactive to the point that droplets freeze homogeneously. By calculating the heterogeneously frozen fraction phet,lab as a sum of droplet volume bins, it is taken into account that freezing of a larger droplet contributes more to the latent heat signal than freezing of a smaller droplet. The probability for a droplet to contain at least one particle depends on its volume. With increasing volume the probability increases that a droplet contains at least one particle. Therefore, the probability for a larger water droplet to freeze heterogeneously is higher than for a smaller water droplet. Hence, freezing of larger droplets dominates the heterogeneous freezing signal and freezing of smaller droplets the homogeneous freezing signal. For example, assuming the latent heat release of the heterogeneous freezing signals to be the same as the latent heat release of the homogeneous freezing signal, the heterogeneous freezing signal represents the freezing of fewer larger droplets and the homogeneous freezing signal the freezing of a high number of smaller droplets.

We consider the natural dust samples investigated here to be representative of atmospheric mineral dust. Most of them were collected from the ground in dust source regions that contribute frequently to the airborne long-range transport. The sieving of the samples with a 32 µm grid rendered them comparable in size with atmospheric mineral dusts, but left the particles otherwise unchanged.

A surprising finding is that the distribution of freezing temperatures of the investigated natural dusts is much more compact than that of the reference minerals for both emulsion measurements (characterizing the typical INP) and bulk measurements (characterizing the best INP). The freezing temperatures Thet,10% of natural dusts with 5 wt % suspension concentration span about 6 K, but those of reference minerals about 12 K (see Fig. 4). Therefore, the reference samples extend far beyond the range covered by the natural dusts in terms of high (as well as low) ice nucleation efficiencies. The dust mixing and/or dust aging (e.g. by coatings) in the natural environment appears to prevent extreme ice nucleation efficiencies (high or low). However, mineralogical composition does matter. Quartz particles were present in all samples (cp. Table 3) of the nine investigated regions (locations 1–9 in Fig. 2). Further, calcite, (Na, Ca)-feldspars (plagioclase), and clay minerals (mainly smectite) were present in seven of the nine regions. K-feldspars (adularia, microcline, orthoclase, and sanidine) were identified still in six regions, however, only the Antarctica, ATD, and Qatar dune samples contained microcline, and the latter only as a minor fraction. These compositions compare very well with the ones of atmospheric mineral dusts compiled by Murray et al. (2012), confirming their assertion that mineral dusts cannot fully account for the high freezing temperatures (above 261 K) observed in the atmosphere. A notable exception might be microcline-containing samples, for which the mineral reference sample microcline N (see Fig. 4) shows bulk freezing even above 270 K. However, out of the 10 investigated samples from dust source regions, microcline was only detected in the Qatar dune sample, and then only as a minor component (4 %). Curiously, this sample did not exhibit a particularly high ice nucleation efficiency (Fig. 3). Boose et al. (2016) found microcline present in one out of four investigated airborne dust samples originating from the Sahara and in three out of eight surface-collected dust samples. If microcline particles are indeed common in atmospheric dusts, they could be relevant for cloud glaciation at temperatures above 260 K, although from the present study this does not seem to be a very likely scenario. In order to investigate this better, atmospheric aerosol samples would need to be analysed with respect to microcline using XRD.

Evaluation of the active particle fraction fact showed that only a part of the mineral dust particles is ice nucleation active. The inactive fraction most probably consists of particles which are composed of inactive minerals such as calcite or muscovite. These particles might at the same time belong to the fraction of particles with smaller size, because larger particles are usually aggregates of different minerals (Reid et al., 2003; Kandler et al., 2011a) and thus more likely contain a mineral that is ice nucleation active. A size dependency is also in accordance with DeMott et al. (2010), who found that the concentrations of INPs active at mixed-phase cloud conditions can be related to the number concentrations of particles larger than 0.5 µm in diameter. Size-selective ice nucleation experiments with pure mineral samples, such as kaolinite, show that larger particles are indeed more effective INPs than smaller ones. This justifies a parameterization of immersion freezing based on particle surface area, as developed by Niemand et al. (2012). The good correlation between mineralogical composition and freezing behaviour suggests that more sophisticated parameterizations should rely on the mineralogical composition based on a source scheme of dust emissions as done by Hoose et al. (2008).

Natural dusts and milled reference minerals were analysed with the objective to investigate whether their ice nucleation activity shows significant differences between different source regions and whether the freezing behaviour can be related to the mineralogical composition. The natural dust samples consisted of calcite, quartz, clay minerals, K-feldspars, and (Na, Ca)-feldspars as major mineralogical components, which is in good agreement with the mineralogical composition of atmospheric mineral dusts. With number distributions that peak for diameters < 1 µm, they are also comparable in size.

The ice nucleation ability of the reference minerals show large variations, much larger than found for the natural dusts investigated in this work. Calcite, dolomite, dolostone, and muscovite seem to induce hardly any freezing. For these minerals, ice nucleation, if present, might not be related to the mineral composition but controlled by impurities. Microcline is able to induce freezing at higher temperatures than all other investigated minerals. In addition, more or less all particles in the two investigated microcline samples are active as INPs. This makes this K-feldspar an exceptionally good ice-nucleating mineral and superior to all other analysed K-feldspars, (Na, Ca)-feldspars, and the clay minerals.

The XRD analysis of the natural dust samples showed that quartz particles were present in dusts from all source regions. (Na, Ca)-feldspars (plagioclase) and clay minerals (mainly smectite) were present in dusts from most source regions. K-feldspars were identified in several source regions; however, only the Qatar dune sample contained microcline as a minor fraction – besides ATD and Antarctica samples, which are not considered to be from typical source regions.

The natural dust samples show very similar freezing temperatures except Antarctica and ATD. These two samples were not collected in source regions of mineral dust aerosols but included as examples of very remote regions or of commonly used reference dust. For all natural dust samples without Antarctica and ATD, 10 % heterogeneously frozen fraction is realized between 244 and 250 K. The 25 % heterogeneously frozen fraction is between 242 and 246 K, and the 50 % heterogeneously frozen fraction is between 239 and 244 K. Bulk freezing occurred between 255 and 265 K. The natural dust samples show active particle fractions fact=0.025–0.32. Taking a potential low bias in our evaluation into account, these numbers might rise to fact=0.1–1. Active particle fractions significantly below 1 are expected considering the share of very small particles and inactive mineral components present in the dusts.

Qualitatively, the mineralogical composition can fully explain the observed freezing behaviour of 5 of the investigated 12 natural dust samples, and partly for 6 samples. Only for the Etosha sample no mineral components could be identified that would explain its high freezing efficiency. This shows that in general the mineralogical composition is a major determinant of the ice nucleation ability of natural mineral dust samples, but cannot entirely explain it. Agglomeration of particles and surface erosion could lead to additional changes in active sites of natural dusts, either enhancing or reducing their efficiency, in comparison to freshly milled reference samples. Comparison of the DSC thermograms of only sieved and additionally milled Antarctica dust shows that milling indeed influences the freezing behaviour and possibly the surface structure of the sample.

The findings of this study are in agreement with previous work that the mineral dusts can hardly account for the glaciation of clouds observed at the highest temperatures in the atmosphere. A notable exception is microcline, for which the temperature of 50 % heterogeneously frozen fraction occurs above 245 K for all concentrations. For the microcline from Namibia, bulk freezing temperatures were even above 270 K. If microcline turns out to be common in atmospheric mineral dusts, this conclusion could be revised. To resolve this question, analysis of airborne dust samples with XRD is necessary, on bulk samples or by EDX on single particles.

While this study shows that mineralogical composition is a major determinant of ice-nucleating ability, in practice most natural samples consist of a mixture of minerals, and this mixture seems to lead to remarkably similar ice nucleation abilities, regardless of their exact composition, so that global models, in a first approximation, may represent mineral dust as a single species with respect to ice nucleation activity. This finding confirms the conclusions by DeMott et al. (2015), who suggested that as a useful first-order approximation in numerical modelling, all mineral dust can be considered as a single type of ice-nucleating particles. However, more sophisticated representations of ice nucleation by mineral dusts should rely on the mineralogical composition based on a source scheme of dust emissions.

To get access to the DSC data and number size distributions, please contact Claudia Marcolli (claudia.marcolli@env.ethz.ch).