The Sphinx Observatory at JFJ is located on a saddle between Mt. Mönch and Mt. Jungfrau in the Swiss Alps (46.330∘ N, 7.590∘ E) at an altitude of 3580 m a.s.l. (see Fig. ). Due to its altitude, the site experiences free tropospheric (FT) air masses, mainly in winter and during nighttime, as well as boundary layer intrusions (BLIs), predominantly in summer and during daytime . The local wind directions measured at the Jungfraujoch are strongly driven by the topography around the Jungfraujoch and are not representative for the larger-scale wind direction at this altitude level (see Fig.  in the Appendix and, for example, ). The principal local wind directions between February and December 2020 were 320∘ (NW) for 62 %, and 150∘ (SE) for 27 % of the time, while calm wind situations below 1 m s-1 had a frequency of 11 %. To the northwest, there is a steep drop of more than 1500 m, whereas to the southeast the elevation decreases steadily over the Aletsch glacier. JFJ has hosted long-term aerosol measurements for more than 30 years and served as a platform for many previous studies to investigate physical, chemical, and optical aerosol properties as well as aerosol–cloud interactions and cloud characteristics see, for example,. There is no appreciable natural source of mineral dust in proximity of the site. Potential local sources of arable dust are isolated agricultural fields 15 km north of JFJ.

In a first step, SDEs were identified using four distinct tracers: the ceilometers in close proximity of JFJ (at Kleine Scheidegg, KSE), the single-scattering albedo Ångström exponent (αSSA) measured at JFJ, dust retrieved from satellite remote sensing generated using the Copernicus Atmosphere Monitoring Service (dustCAMS), and FLEXPART particle residence times over the Saharan region. The methodology of all tracers is introduced in the upcoming dedicated subsections. If all tracers show a signal, the event is classified as high-confidence Saharan dust event (hcSDE). If at least one of the tracers shows a signal but not all tracers agree, then the event is classified as a low-confidence Saharan dust event (lcSDE). The beginning and end of each SDE reflect the beginning and end of the onset of the earliest tracer or the decay of the latest tracer, respectively. If none of the tracers shows a signal, the event is classified as non-Saharan dust event (non-SDE). Only after the classification of SDEs are the INP measurements at JFJ analyzed.

The INP concentrations were measured with an autonomous continuous-flow diffusion chamber, the automated Horizontal Ice Nucleation Chamber HINC-Auto;, at T=243.15 K and Sw=1.04 at the Sphinx Observatory on JFJ. All INP concentrations are in units of INP std L-1 (per standard liter, normalized to T=273.15 K and an atmospheric pressure of p=1013.25 hPa). Ambient air is sampled using a heated total aerosol inlet 293.15 K;, which also feeds other aerosol measurements (a condensation particle counter (CPC), a scanning mobility particle sizer (SMPS), an optical aerosol spectrometer (OAS, FIDAS 200, Palas GmbH, Germany), two nephelometers, two filter-based absorption photometers, and a particle mass monitor). Due to the low-volume flow of HINC-Auto and the sampling line geometry, particle survival measurements using an optical particle counter (MetOne GT-526S) and an OAS showed limitations when sampling particles d>2.5 µm (see Appendix for detailed information), despite the fact that no impactor was used upstream of HINC-Auto. According to particle size distribution measurements from the OAS, particles with d>2.5 µm contributed during the investigated SDE periods on average to 35 % to the overall particle surface area (see Fig.  for an example of particle size distributions during an SDE). This is a noteworthy limitation, as studies have reported a majority of INPs at T>248 K to be supermicron particles e.g.,; for a constant ice-active fraction, INP concentrations scale with total particle surface area for a given temperature e.g.,. However, instrument comparisons typically report substantially larger discrepancies between individual instruments than 35 % e.g., see.

INPs are detected if particles with an optical diameter of ≥4.0 µm are counted at the chamber exit. Only particles with a diameter < 2.5 µm are sampled by the chamber, and the maximum expected size of a droplet that activated on a 2.5 µm particle is well below 4 µm. Therefore, the method is robust as only ice crystals grow past the set size threshold of 4 µm in the set conditions. The frequency of the INP measurements is every 20 min (15 min sampling plus 5 min background), corresponding to N=19561 measurements between 7 February and 31 December 2020. See for more information on the sampling and derivation of the INP concentrations. During the investigated time period, construction work was undergoing in the tunnels of JFJ, as described in . In case of local pollution, high-frequency fluctuations in particle concentrations were observed (e.g., in the CPC measurements). To filter periods of local pollution, the following two methods were applied. Firstly, the smallest optical particle counter (OPC) size bin of HINC-Auto (0.3 µm) was analyzed with regard to fluctuations. A typical INP measurement includes 180 OPC sequences of 5 s each. For each 5 s sequence, the cumulative number of particles with a optical size of ≥0.3 µm is counted. If the count of a 5 s sequence is 30 % higher than the average count of all 180 sequences, the corresponding INP measurement is flagged. Secondly, fluctuations in the CPC measurements at JFJ were analyzed. If the 1 min average concentration changed by more than 40 particles cm-3 from the previous 1 min average, the current and the subsequent 5 min of measurements were flagged. These thresholds were chosen based on a qualitative assessment of the fluctuations in the OPC and CPC measurements. All INP measurements containing flagged CPC periods were flagged as well. Applying both methods is estimated to conservatively flag potentially polluted INP measurements and leads to Nunflagged=14216 (72 %) measurements unaffected by local pollution.

False-positive counts can arise in HINC-Auto. Frost grows on the chamber walls, breaks off, and is detected at the outlet as ice. This happens irrespective of whether ambient or particle-free air is sampled with HINC-Auto. Thus, to correct the measured INP concentrations, the number of frost particles is measured separately and subtracted from the uncorrected INP measurements. This is done by sampling particle-free air for a period of 5 min before and after an ambient air measurement. During these periods, the number of false-positive frost particles is counted and time-proportional subtracted from the ambient air measurement in between. The recorded false-positive counts per unit time follow a Poisson distribution. Therefore, in a fraction of cases more or fewer false-positive counts per unit time are recorded during the particle-free measurements than during the ambient air measurement. This results in fluctuations of the measured INP concentrations even if the true atmospheric INP number concentration were to remain constant. The standard deviation of the resulting probability density function corresponds to the stated counting uncertainty of ± 1σ, which is provided with INP concentrations stated in the present work. This counting uncertainty is also considered to be the 1σ limit of detection (1σ LOD) for a single data point. However, in the present work all background-corrected INP concentrations are retained, including positive values below the 1σ LOD and negative values. This approach ensures that the random noise in background-corrected INP values caused by subtracting the mean frost particle counts does not introduce a systematic bias in mean or median values, which would occur if data below the 1σ LOD were discarded.

The decay of radon-222 (222Rn) is measured at JFJ according to . 222Rn is a tracer of BLI at high-altitude sites, as there are no substantial sources of 222Rn in the FT. 222Rn is formed by the decay of naturally occurring radioisotopes in minerals. If the surrounding rock at a mountain site is covered by snow and ice, as is mostly the case at JFJ, the formed 222Rn is inhibited from mixing into the ambient air, leading to low 222Rn concentrations when the site is experiencing FT air masses. On the other hand, 222Rn is released relatively homogeneously in time and space from soils and gets well-mixed within the PBL, leading to higher 222Rn concentrations at JFJ during periods of BLI from air with contact to bare soil. Further information can be found in .

When looking at remote sensing data of the PBL over alpine terrain e.g.,, the air masses present at the site are not binary (either FT or BLI), but the PBL can gradually get mixed with FT air . Consequently, setting a binary threshold does not resemble an atmospheric process or represent the recorded data. Therefore, we use an approach where each 222Rn concentration corresponds to a probability of the sampled air to be of free tropospheric origin (as explained below). For this, the 222Rn concentrations at JFJ were analyzed between 1 January 2009 and 31 December 2020, as shown in Fig. . Founded on the theory where the concentration of pollutants in the atmosphere follows a lognormal distribution , two normal distributions were fitted so that the sum of them closely reproduces the observed bimodal distribution of the logarithm of the 222Rn concentrations. One distribution corresponds to 222Rn concentrations found in the FT (fit parameter: μFT=-0.139 Bq/std m3, σFT=0.239 Bq/std m3) and the other distribution to the concentrations measured during BLI (fit parameter: μBLI=0.403 Bq/std m3, σBLI=0.238 Bq/std m3).

Besides 222Rn, the total number concentration of particles with diameters larger than d≥90 nm (N90) is another tracer for BLI . The threshold above which the air mass is likely boundary layer intruded is N90≥ 100–150 cm-3, depending on the season . Ceilometer measurements can also be used to infer the top of the continuous aerosol layer, which indicates the top of the planetary boundary layer and, subsequently, whether JFJ is within the PBL. However, this method has not been used in the present study. A further tracer of BLI is the ratio of total reactive nitrogen (NOy) to carbon monoxide (CO) according to . Both NOy and CO are emitted from anthropogenic sources; however, NOy reacts and decays on the order of days, while CO can be considered inert within the same time period. Consequently, the ratio of NOy to CO decreases with increasing aging time, leading to smaller ratios found in the FT compared to BLI . CO is continuously measured with a cavity ring-down spectrometer at JFJ . NOy was measured until March 2020 with a chemiluminescence detector after conversion to NO on a heated gold catalyst (573 K) in the presence of CO as a reducing agent . An analysis of 222Rn, N90, and NOy / CO data collected at JFJ between 1 January 2017 and 31 December 2018 showed for N90≥100 cm-3 to be indicative of BLI, but concentrations of N90< 100 cm-3 to not always be indicative for FT air masses, as indicated by 222Rn and the NOy / CO ratio (see Fig.  in the Appendix). Therefore, in the present study a sigmoid function forces periods with concentrations of N90 above a smooth threshold (to cover the seasonal variations of the threshold) to have a low probability to be sampled within the FT, even if the 222Rn criterion indicates FT, but allows for periods with concentrations below the smooth threshold to be either FT or BLI depending on the 222Rn concentration. The threshold midpoint (N90,TH) and smoothing, using the slope factor k (here N90,TH=120 cm-3 and k=0.1), have been chosen to best agree with season-specific thresholds found in the literature . The probability of the sampled air to be of free tropospheric origin (PFT) is according to Eq. (): 1PFT=PDFFT(222Rn)PDFFT(222Rn)+PDFBLI(222Rn)11+ek(N90-N90,TH), where PDF is the probability density function with the FT or BLI fit parameters, respectively, according to 2PDFFT(222Rn)=1σFT2πe-12log⁡10222Rn-μFTσFT2.

In the present study, NOy / CO was not used because NOy measurements were discontinued at JFJ in March 2020.

Figure  shows an example of an SDE with all introduced tracers, the air mass type, and the INP concentration. In Fig. b, the attenuated backscatter of the ceilometer at KSE showed an increase from background levels to ∼1 m-1 sr-1 at altitudes similar to JFJ after midnight on 10 July, indicating the presence of an aerosol plume. At 11:00 UTC, the signal of the aerosol plume was attenuated by the low-level clouds during the remaining period of the plume event. αSSA decreased to below zero with decreasing wavelength after 02:00 UTC on 10 July (Fig. c). The signal becomes less separated with αSSA above zero at 15:00 UTC and noisy after midnight of 11 July, indicating the end of the SDE according to this tracer. dustCAMS mass concentrations exceeded the threshold concentrations on 9 July at 23:00 UTC, as shown in Fig. d, peaking at 03:00 UTC of 10 July with 19.6 µg m-3, followed by a decay, until falling below the threshold at 20:00 UTC. FLEXPART particle surface residence times in Fig. e indicate between 03:00 UTC on 10 July and 05:00 UTC on 11 July that the air mass is expected to have had ground contact over the Saharan domain. Following all four tracers showing a signal, the SDE was classified as hcSDE. The INP concentrations show an increase from background INP concentrations on 9 July at 23:00 UTC (Fig. a) to concentrations above 200 INP std L-1, followed by lower concentrations coinciding with αSSA relaxing back to values above zero for 1 h. After a brief, second increase in INP concentrations, a decline to background levels at midnight of 7 July followed. No tracer shows an identical onset and decline as observed in the INP concentrations; however, αSSA was the closest. The BLI air masses were present during the SDE, with 222Rn initially at FT levels at 04:00 UTC on 10 July, followed by a rapid change to BLI levels. At the end of the SDE, N90 indicated FT conditions, while 222Rn still pointed to BLI, resulting in the air mass being classified as BLI.

The mass concentrations of particulate matter with aerodynamic diameters below 2.5 µm (PM2.5) and below 10 µm (PM10) are continuously recorded with a white light optical aerosol spectrometer (Fidas 200, Palas GmbH, Germany). The data are provided with a temporal resolution of 10 min, with each data point reflecting the average particulate matter over the past 10 min. The spectrometer measures the intensity of single particles at an angle between 85 and 95∘ and infers the particle diameter solving the inverse Mie problem. The sensitivity of the spectrometer is checked six times per year with calibration dust.

To assess the atmospheric transport of dust and the presence and retrieved phase of nearby clouds in the case study (see Sect. ), post-processed lidar data from the Cloud-Aerosol LIdar with Orthogonal Polarization (CALIOP) instruments of the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellites was used. CALIPSO is part of the NASA A-Train constellation , which was launched in 2006 to a sun-synchronous polar orbit with 98.2∘ inclination at an altitude of 705 km . The data were accessed via https://www-calipso.larc.nasa.gov/products/lidar/browse_images/exp_index.php?d=2020 (last access: 10 April 2021).

Figure  shows box plots of the measured INP concentrations for all of the 26 SDEs, classified either as hcSDE or as lcSDE, compared to the box plot of measured INP concentrations during all non-SDE periods. Every SDE shows elevated levels of INP concentrations compared to the non-SDE periods, rendering every SDE to carry a higher INP loading than what is present during non-SDE periods. However, the concentration of INPs varies across SDEs, as the SDE with the lowest INP concentration (median =1.7 INP std L-1) is close to the median of background INP concentration and 2 orders of magnitude smaller than the median of the SDE with the highest INP concentration (median = 161 INP std L-1). Three SDEs (12 % of all SDEs) showed median INP concentrations above 100 INP std L-1.

The 12 lcSDEs accounted for 33 % of the total SDE duration and showed a lower median INP concentration of 13.1 INP std L-1 compared to 28.5 INP std L-1 for the 14 hcSDEs. This reflects that lcSDEs have the tendency of showing lower INP concentrations than hcSDEs, yet the SDE with the lowest 95th percentile of 7.1 INP std L-1 was a hcSDE in May, while the SDE with the highest 95th percentile of 881.1 INP std L-1 was an lcSDE in June. Therefore, increased INP concentrations can be expected if at least one of the mentioned four tracers shows an SDE signal. Furthermore, all tracers showing a signal is not indicative for highest INP concentrations to be expected. As the αSSA is the tracer with least correspondence to the other SDE tracers, this is specially relevant for sites where the presence of dust events is inferred from the αSSA alone, as has been done in previous studies at JFJ .

Figure a shows the histogram of observed INP concentrations. The observed concentrations during SDEs are responsible for all INP concentrations above 102 INP std L-1. If a lognormal curve is fitted to all INP concentrations (blue dashed line in Fig. a; fit parameter: μ=100.517 INP std L-1=3.3 INP std L-1 and σ=100.516 INP std L-1=3.3 INP std L-1), both tails of the measurements (black line) are symmetric and above the fitted curve. Therefore, the frequency distribution of all INP concentrations shows a leptokurtic lognormal distribution, which is consistent with the theory by and other studies e.g.,. If the fitted curve is compared to the frequency distribution of all non-SDE INP concentrations (pink line), the non-SDE INP concentrations show a distinctly less symmetric lognormal distribution, as the high concentration tail above 10 INP std L-1 falls below the fitted curve, while the low concentration tail remains above the fitted curve. This could mean that the kind of INPs detected during SDE also contributes to the overall INP population during non-SDE periods, but chemical analysis would be necessary to categorically conclude this. After all, SDEs are operationally defined periods, where dust concentrations become so dominant that their signature can be more clearly measured, and the properties of dust particles outweigh competing effects from other aerosol species (e.g., absorption, back- and forward-scattering characteristics).

HINC-Auto samples an approximately constant volume flow of ambient air, and counts the number of INPs after exposure to the set temperature and supersaturated saturation ratio within the chamber. The ratio of INPs detected during all SDEs to all INPs detected during the entire analyzed period is calculated to quantify the contribution of Saharan dust, including a 1σ limit of detection (LOD) to express the uncertainty involved:

3∑INP during SDEs±LOD∑total INP±LOD=90938±90.4INP121691±265.1INP=74.7±0.2%.

Thus, 74.7 ± 0.2 % of all INPs were detected during SDEs. Whether this is because of the ice nucleation activity of the mineral dust or because of biogenic material content, as proposed by other studies e.g., see, is outside of the scope of this study. Biogenic material on the dust particles as the predominant cause of the ice nucleation activity of the particles would have two main implications. Firstly, the underlying mechanism leading to ice nucleation might differ, as proteins and other macromolecules could induce the ice nucleation as compared to topological mineral surface features, such as cracks and pores. Secondly, it would raise the question of the source of the dust particles containing ice-active biogenic material to be potentially from dried lake beds in desert regions or agricultural regions that are not differentiated in this work. In addition, a contribution of arable dust cannot be ruled out; however, during the 26 SDEs the modeled FLEXPART particle surface residence times indicated in 24 cases that the air mass had robust surface contact in the Sahara, and in the other two cases, they had weak surface contact. The median INP concentration during SDE periods with FT air masses is 17.3 INP std L-1, while with BLI it is 23.7 INP std L-1. If we assume arable dust to show a substantially larger signal within the well-mixed PBL than in the FT, we could attribute the difference between median INP concentrations in FT SDE periods and BLI SDE periods to be because of arable dust, which is 6.4 INP std L-1. There were three SDEs detected with lower concentrations than 6.4 INP std L-1: one with a signal in αSSA, one without, and one where the nephelometer was offline. If during SDE periods the median contribution of arable dust was 6.4 INP std L-1, then this should be also similar during non-SDE periods. However, the median INP concentration during non-SDE periods was only 1.1 INP std L-1. We do not see any indications of why during SDEs the contribution of arable dust should be substantially larger than during non-SDE periods. If at all, we expect arable dust would contribute to BLI INP concentrations during non-SDE periods. It is fair to assume for background INPs present during non-SDE to be also present during SDE. Assuming a constant concentration over time, background INPs would contribute as little as 0.5 % to the INP population during SDEs.

Since SDEs consist only of the high concentration tail of the lognormal distribution of dust in the atmosphere after successive random dilution by large- and small-scale dynamics, also visible in the dustCAMS frequency distribution in Fig. b, it is likely that ice-active dust is also present outside of SDE periods but in smaller concentrations. For the following assessment, we assume that a constant mass fraction of all dust carried to JFJ will act as INPs. This is a bold simplification; for example, found INP concentrations to vary within a factor of 7 for the same dust mass concentration; however, it provides a rough estimate of the contribution of dust to the INP concentration at T=243 K and Sw=1.04 at JFJ. The variable dustCAMS will be used as a proxy for the ambient dust mass concentration. The ratio of dustCAMS advected to JFJ during all SDEs to all advected dust to JFJ during the entire analyzed period is 4∑dustCAMS during SDEs∑totaldustCAMS=9538µg12466µg=76.5%.

A literature search did not yield information about the uncertainty of dustCAMS. If 76.5 % of the dust is responsible for 74.2 ± 0.2 % of the INPs (74.7 ± 0.2 %–0.5 %) and 23.5 % of all dustCAMS was advected to JFJ during non-SDE periods, we estimate that about 23 % of the INPs measured during non-SDE periods were dust-related with our assumption that a constant mass fraction of dust acts as INPs. Therefore, the total contribution of dust to the INP population measured at JFJ at T=243 K and saturation ratio of Sw=1.04 is estimated to be 74 % + 23 % ≈ 97 %. Note, during non-SDE periods, dust contributed 23/25.3 % ≈ 91 % to the overall INP population. To validate the stated contribution in future studies or investigate the presence of biogenic material causing the ice activity, we propose to separate INPs from the bulk aerosol population to analyze the chemical composition of the INPs as well as to study the surface using scanning electron microscopy. However, to our knowledge, such equipment to separate INPs has not been used in a continuous annual study. Detailed suggestions of how this can be achieved in a long-term automated study by modifying HINC-Auto is presented elsewhere and is beyond the scope of the current article.

We would like to emphasize that there is substantial uncertainty involved in the stated fraction, as the assumption of a constant ice-active mass fraction in all dust particles changes with dust type e.g.,and references therein. Also, the uncertainty of dustCAMS at low concentrations is expected to be significant but remains unquantified. Furthermore, the statement is only applicable for the stated location, temperature of T=243 K and saturation ratio of Sw=1.04. At warmer temperatures mineral dust shows a lower ice activity e.g.,and references therein and subsequently plays a lesser role at JFJ such that other species become the dominant population. This is consistent with observations by , where INPs active at 265 K showed a correlation with ambient temperature and during SDEs were on the same order of magnitude as during non-SDE periods. This is expected, as at 265 K the dominant type of INP is different than at 243 K (e.g., biological vs. mineral dust). Also, at measurement locations further away from the dust source than JFJ is or locations closer to local sources in the PBL, the contribution of dust on the total INP population can be expected to be much smaller.

This raises the question of whether the INP concentrations at JFJ can be inferred meaningfully from dustCAMS. Figure  shows the dustCAMS mass concentration at the time of every INP measurement with the corresponding INP concentrations and PM10 concentrations. The Spearman's rank correlation coefficient of ρ=0.509 indicates that there is a correlation between the dustCAMS and INP concentrations; however, the R2=0.276 indicates there is only a weak linear correlation. This does not come as a surprise, as the INP concentration is a particle number concentration per volume of air that tends to scale with particle surface area (∝r2) for an identical INP type or air mass dominated by a certain INP species or with the number concentration of viable dust particles, while dustCAMS provides a mass concentration (∝r3). Given that the distribution of the dust particles during the SDEs is not lognormal (see Fig. b) suggests that the size of dust particles varies; thus, the number concentrations will not scale with dust mass. Therefore, dustCAMS is also compared to PM10, which both are in units of mass per volume of air. Interestingly, with ρ=0.543 and R2=0.273 the pattern does not change much, indicating the level of uncertainty within dustCAMS. A comparison of the observed PM10 with the PM10 from CAMS showed a similar agreement with ρ = 0.491 and R2=0.265 and better agreement with ρ = 0.591 and R2=0.336 if only concentrations above 1 µg m-3 are considered. As the relative measurement uncertainty in lower INP concentrations and presumably also in low dustCAMS concentrations increases, a higher degree of correlation would be expected at high concentrations, which is, however, not the case. This analysis strengthens the caveats of our earlier conclusion regarding the dust contribution to INP concentrations at JFJ.

Figure a shows the INP concentrations measured during a classified SDE in spring, peaking at 419 INP std L-1. Figure b shows the ceilometer attenuated backscatter signal at KSE, with the increased attenuated backscatter due to the dust particles up to 5000 m a.s.l., and in Fig. c the attenuated backscatter evaluated at similar altitudes as JFJ, with filtered ceilometer measurements when a cloud was below or at the same altitude as JFJ. Similar altitudes at KSE rather than the exact altitude of JFJ were chosen as, depending on the wind direction, orographic lifting or subsidence is to be expected between KSE and JFJ. Furthermore, eddies forming on the leeward side of the mountain ridge can induce mixing of air masses from higher or lower altitude to JFJ, depending on wind speed, wind direction, and atmospheric stability. The similar trends in observed INP concentrations and attenuated backscatter are not just a feature of the SDE on 16–17 April but across all investigated data where the ceilometer signal is not attenuated by clouds or precipitation. Another example is provided in Fig. , where it becomes apparent that the observed INP concentration does not always best agree with the backscatter signal retrieved at the altitude of JFJ but sometimes rather with the signal retrieved at a lower or higher altitude. This also changes over time as the wind situation changes throughout the day e.g., see.

Figure  shows the observed relation between the INP concentration and the attenuated backscatter of three classified dust events. With Spearman's rank correlation coefficients ranging between 0.517≤ρ≤0.815, there is a correlation between the two quantities. As both INP concentration and the backscatter coefficient scale to a first approximation with the cumulative surface of the ambient particles, a linear fit can be applied. If a linear fit is applied to the logarithm of both quantities, including all altitude levels, the slopes of the fitted curves are very similar, but the offsets differ, resulting in a factor of up to 4.5 different INP concentrations for the same attenuated backscatter. Considering the limiting factors, like the 4.4 km distance between the measuring sites; the lifting and subsidence depending on the synoptic situation, modified by the orography; and the analysis of the attenuated backscatter signal rather than the corrected backscatter signal, we consider the agreement such that it allows us to use the signals of the ceilometer network to study the spatiotemporal evolution of INP concentrations and investigate their atmospheric pathways. For this purpose, a fit was computed for all data (SDE + non-SDE) where the ceilometer signal is not attenuated by clouds or precipitation (ρ=0.63):

5INP243K conc.=102.062log⁡10(ABSc)+2.81, where INP243K conc. is the INP concentration at T=243 K and Sw=1.04 in INP std L-1, and ABSc is the attenuated backscatter (in m-1 sr-1). We propose the fit in future work to be validated at other locations where ceilometer data are available. As previously shown, the INP concentrations at JFJ are dominated by mineral dust particles, which renders the obtained fit likely to only be valid for when mineral dust is the dominant species amongst the INP population.

The ceilometer at KSE often showed upstream virga prior to dust plumes arriving at JFJ. These virga are starting at an altitude above 5000 m a.s.l. and appear to be connected to dust plumes (see Fig. b). To illustrate this, two dust events were analyzed in more detail. Figure a shows a map of the dust concentration from CAMS on 9 February 2020, 06:00 UTC, with the location of three ceilometer stations: Freiburg (FRE) in southern Germany, KSE, and Saint-Auban (StA) in the south of France. An anticyclone with a surface pressure of 1030 hPa at sea level centered over Serbia extended to the Alps, while a weak cold front with core above Iceland passed Germany. Figure b correspondingly shows the CAMS dust concentration on 16 April 2020, 06:00 UTC, with the ceilometer stations Deuselbach (DEU) in the western part of Germany and KSE, where an anticyclone with a center over Croatia extended to central France and Germany with a surface pressure of 1020 hPa at sea level.

Figure  shows the ceilometers StA, KSE, and FRE between 7–10 February. On 8 February, at 23:29 UTC, INP concentrations exceeded background concentrations, indicating the onset of an SDE, and it peaked on 9 February at 02:59 UTC with 42 INP std L-1. The corresponding signal can be seen in the attenuated backscatter at KSE during the same period, initially at an altitude between 3500 and 4600 m a.s.l. and descending over time. A faint signature of the plume is visible before 8 February, 23:30 UTC, at the altitude range between 4500 and 5500 m a.s.l. and going back until 8 February, 13:00 UTC, where it connects to signatures from sedimenting hydrometeors with equal upper and lower limits. These hydrometeors in turn appear to have sedimented from 11 400 m a.s.l., forming a virga. Analogous behaviors in the signals of the plume with each corresponding virga are also visible in the ceilometers StA and FRE (see Fig. ); however, the signature of the plume is more distinct. Radio soundings at Payerne, 84 km northwest of KSE, observed on 8 February, at 00:00 UTC at 11 400 m a.s.l., with a temperature of 217 K and saturation ratio with respect to liquid water and ice of Sw=0.39 and Si=0.68, respectively, were well below the saturation ratio needed for the homogeneous nucleation of solution droplets of ∼Si=1.52 at 217 K . The tropopause was above 11 650 m a.s.l. according to radio soundings. Ice saturation was detected between 8850 and 8880 m a.s.l. and between 10 200 and 11 200 m a.s.l. Below 7790 m a.s.l., the saturation dropped from Si≥0.8 to Si≤0.15. This raises the hypothesis of whether the ice-active particles nucleated the ice clouds at altitudes above 7790 m a.s.l., with subsequent sedimentation of the ice crystals to below 7790 m a.s.l., where they sublimated leaving behind dry INPs. These INPs further sedimented, however, due to the lower mass, at a much smaller rate. Thus, signals of the plume appear, in the ceilometer, more elongated after sublimation (e.g., in FRE at 13:00 UTC on 8 February) compared to prior in the virga. What opposes this hypothesis is the fact, that no Saharan dust plume can be observed in the ceilometer measurements at altitudes above 7790 m a.s.l.

Figure  shows another example of the observed pattern, with the ceilometers KSE and DEU. The simultaneous ceilometer and INP measurements at KSE and at JFJ, respectively, which are shown in Fig. a, indicate that the Saharan dust plume on 16–17 April contains INP number concentrations above 100 INP std L-1; however, no connection between the Saharan dust plume and a virga is apparent. The Saharan dust plume can be tracked north to other ceilometer locations, such as DEU, following the outline of the dust plume in Fig. b. At DEU, shown in Fig. b, the lower end of the virga and the onset of the dust plume are colocated as before in the case study from 8 February. However, also in this case no signal of the Saharan dust plume is apparent above 5500 m a.s.l. The sounding at Idar-Oberstein from the Deutscher Wetterdienst (15 km east of DEU) reported a tropopause at 10 200 m a.s.l. and ice supersaturation between 8370 and 10 370 m a.s.l. Below 8370 m a.s.l., there was a steady decrease from ice saturation to Si=0.17 at 5100 m a.s.l. A three-dimensional kinematic backward trajectories analysis was carried out using LAGRANTO , with wind fields from the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System (IFS) HRES model with a horizontal resolution of 0.1∘ × 0.1∘. The analysis showed trajectories to originate from the Sahara within Algeria, where they were close to the surface (≤300 m above ground) on 14 April, followed by an ascent to 300 hPa (∼9000 m a.s.l.) above KSE. Observations from the CALIOP lidar aboard the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellite show large amounts of dust over Algeria up to 6000 m a.s.l. on 14 April. On 15 April, CALIPSO did not pass over the area west of Sardinia, where the backward trajectories would predict the location of the air mass containing the Saharan dust. On 16 April at 12:30 UTC, CALIOP retrieved an aerosol signature between 11 000 and 13 000 m a.s.l. over central Italy (43.3∘ N, 12.7∘ E), surrounded by cirrus clouds, but it could not identify the type of aerosol. Assessing all presented data, the hypothesized pathway of mineral dust being lifted to cirrus altitudes – where it nucleates ice to form cirrus clouds and sediments within the ice crystals to lower altitudes, where after ice sublimation the dust is left as residuals – is plausible yet remains to be proven at this point. We plan on addressing the hypothesis in future work. Such a transport pathway poses some relevant implications for atmospheric processes. First, the simulated backward trajectory analysis will not well reproduce the history and thus the origin of the particles, as models backtrack air parcels and not particles, which are subject to varying settling velocities when ice is nucleating and growing on them. And secondly, after sublimation of the surrounding ice, the residual mineral dust particles can be pre-activated, e.g., by retaining pore ice due to the inverse Kelvin effect . If the pre-activated particles are again exposed to ice supersaturated conditions, spontaneous ice crystal growth can occur at much lower Si than with other INPs . To capture and study this effect, design adaptations need to be implemented in our sampling equipment in upcoming studies, as the pore ice sublimates or melts in the heated and dried sampling lines upstream of the INP counter used here. We attempted such a measurement of the number concentration of pre-activated particles in this way, using a supercooled diffusion dryer at JFJ before directly sampling the ambient particles with HINC-Auto, in addition to sampling with a reference chamber with the standard heated inlet. Nevertheless, we were unable to reach any reliable conclusions. In part, due to the clogging of the inlet from frost build-up, the conditions along the trajectory of the potentially pre-activated particles had to be precisely adjusted, such that neither particle sedimentation due to activation nor pore ice sublimation occurs. Furthermore, the set conditions do not affect all pore sizes similarly, and pore ice can melt if the pore ice shrinks below the critical size of an ice embryo within the pore with a constrained width. In summary, there were too many parameters that needed to be simultaneously and meticulously controlled.