The chemistry transport model used in this study is the MUltiScale Chemistry Aerosol Transport (MUSCAT) coupled online with the COnsortium for Small-scale MOdelling (COSMO) v5.05 model. COSMO, developed by the German Weather Service (Deutscher Wetterdienst, DWD), is a non-hydrostatic regional weather prediction model that solves the fundamental equations of atmospheric dynamics on a terrain-following grid . MUSCAT is the online-coupled chemistry transport component, computing the atmospheric transport of aerosols through time-dependent mass balance equations driven by COSMO meteorological fields . In this work, mineral dust aerosols are represented as passive tracers, i.e., they are not subject to chemical aging or chemically reactive transformations.

The coupled atmosphere-aerosol model system, COSMO-MUSCAT, has been widely applied and evaluated for Saharan dust studies. Previous validation efforts have demonstrated its capability to reproduce dust source activation, transatlantic transport, and regional dust transport under different meteorological conditions . The specific configuration used here has been further evaluated against atmospheric dust loading observations for the North Africa region in and .

The atmospheric life cycle of dust aerosols in MUSCAT is represented through a set of physical parameterizations dynamically coupled to COSMO meteorology and updated at every advection step. The main processes include: (1) dust emission, parameterized following with modifications to incorporate mineralogical soil fractions as described in ; (2) aerosol transport, solved using a third-order upwind advection scheme with time-splitting integration ; and (3) aerosol deposition, accounting for both dry and wet removal processes. Dry deposition is parameterized following the formulations of and , while wet deposition (including in-cloud scavenging or rainout, and below-cloud scavenging or washout) follows the approaches of and , with detailed implementation described in .

Dust emission is a non-linear process initiated when near-surface wind velocity generate sufficient vertical shear stress at the soil surface to initiate particle mobilization. In COSMO-MUSCAT, threshold friction velocities are calculated following the parameterization of , which accounts for size-resolved soil particle mobilization and depends on soil texture, surface roughness, vegetation, and soil moisture.

Emission fluxes are then computed iteratively using the scheme of , first implemented in COSMO-MUSCAT by . Fluxes scale with the cube of the wind friction velocity, derived from COSMO-simulated near-surface winds, and are further modulated by vegetation cover and soil moisture. Particle uplift occurs when the effective friction velocity exceeds the size-dependent threshold (Ut*), which is controlled by the erodible particle diameter (Dp), the aerodynamic roughness length of the total surface (Z0), and the smoother, erodible fraction roughness length (z0s).

To account for the sheltering effect of roughness elements, a drag partitioning approach is utilized. This method defines an effective friction velocity (Ut*), representing the total friction velocity required to mobilize size-dependent particles on a rough surface. It is related to the smooth-surface threshold friction velocity (Uts*), which is the threshold required for mobilizing the same particles under idealized, flat, conditions, via: 1Ut*(Dp,Z0,z0s)=Uts*(Dp)feff(Z0,z0s), where feff is the effective drag partition factor (ranging from 0–1). This factor accounts for the reduction in wind erosive potential due to surface roughness and it is calculated following as: 2feff(Z0,z0s)=1-ln⁡Z0z0s/ln⁡0.3510z0s0.8.

For bare desert surfaces, the aerodynamic roughness length (Z0) is prescribed as 0.001 cm, following the recommendation of , while z0s is obtained from global satellite-derived dataset . The value of Uts* is calculated as a function of particle diameter (Dp), particle density, and air density (ρa), representing a balance between aerodynamic and gravitational forces.

The actual surface friction velocity exerted by the atmosphere (U*) is estimated from COSMO's 10 m wind speed components (u10, v10) and air density. These are combined to determine an effective wind speed at height Z, denoted as U(Z). Assuming adiabatic conditions and a logarithmic wind profile within the boundary layer, U* is calculated as: 3U*=U(Z)κln⁡ZZ0 where κ=0.4 is the Von Karman's constant, and Z represents the height of the first model layer.

Dust emission is triggered when this surface friction velocity exceeds the size-dependent threshold friction velocity (U*≥Ut*). This process is further modulated by local soil conditions such as moisture, which is accounted by an additional coefficient in the threshold calculation, following .

Once the threshold is surpassed, the emitted flux is represented as the horizontal particle flux (G), which scales with the cube of the friction velocity and accounts for the soil particle size distribution: 4G=ρag⋅U*3⋅∑i1+Ut*(Dpi,Z0,z0s)U*×1-Ut*2(Dpi,Z0,z0s)U*2⋅Brel-iforU*≥Ut*, where g is gravitational acceleration, and Brel-i represents the relative basal surface area of size fraction i. In MUSCAT, the soil is represented by 196 discrete size fractions.

Because the threshold friction velocity does not scale linearly with particle diameter, an accurate representation of the soil size distribution is required. In the parametrization, the soil particle mass size distribution (PSD) is modeled using a multi-modal log-normal distribution: 5dm(Dp)dln⁡(Dp)=∑j=1nmj2πln⁡(σj)exp⁡(ln⁡Dp-ln⁡MMDj)2-2ln⁡2σ. where j denotes the size mode (clay, silt, sand in the present MUSCAT setup) mj is the mass fraction of mode j, σ is the geometric standard deviation (set to 2.0 independent of the mode), and MMDj are the mass median diameters, set to 2.0 µm (clay), 15.0 µm (silt), and 150.0 µm (sand), respectively.

To determine the basal (projected) surface area distribution, the PSD is transformed assuming spherical particles of uniform density: 6dBt(Dpi)=dm(Dpi)23ρDpi.

The total basal surface area (Bt) is obtained by integrating Eq. () across all particle sizes. Normalizing dBt(Dpi) by Bt yields the relative basal surface area distribution Brel-i, which is used in Eq. () to allocate the horizontal flux across particle sizes.

Saltation and particle bombardment processes are incorporated by iteratively adjusting Brel-i for each size class, accounting for the momentum transfer from saltating particles to finer dust grains. Following this adjustment, the total horizontal flux is computed.

The fraction of the horizontal flux that becomes airborne is parameterized as a vertical dust flux (F): 7F=ω⋅Aeff⋅G⋅(1-Asnow)⋅Iθ, where ω is the sandblasting efficiency, prescribed per soil type based on the local clay, silt, and sand fractions . Aeff denotes the erodible surface area, Asnow is the snow-covered fraction, and Iθ represents a soil moisture correction following . The vertical flux is then partitioned into five transport-relevant size bins (Table ) to generate size-resolved dust fluxes for atmospheric injection.

The overall magnitude and spatial variability of emissions are determined by the interplay of these surface conditions. The influence of vegetation cover on Aeff is threshold-dependent. In desert regions, dust emission is inhibited once fractional vegetation cover exceeds 0.5; below this threshold, the erodible area is scaled linearly to account for the partial sheltering of the surface . Similarly, Iθ reduces emissions linearly once soil moisture exceeds a critical threshold determined. While snow cover also inhibits dust emissions, its parametrization is deactivated in this model configuration given its negligible influence over the Sahara Desert. Together with the particle size distribution and surface roughness, these factors control the efficiency and spatial heterogeneity of dust emissions in the model.

In the “original” approach by , mineralogical fractions are prescribed by directly mapping the mineral soil particle size distribution (PSD) from GMINER onto the mineral-resolved aerosol PSD as represented by the upper panel of Fig. . A key limitation of this approach is that the modifications to the bulk PSD caused by the emission process do not affect the mineral-resolved PSD. This overlooks the fact that the emission process explicitly alters the overall dust PSD, as predicted by dust emission theories .

This inconsistency stems from the structure of GMINER in combination with 's emission scheme, since GMINER provides mineralogical mass fractions only for the finest soil classes (silt and clay). These are assumed to represent the airborne dust range, given their capacity to remain suspended for several days. However, the emission scheme of (Eqs. –) requires the full soil PSD to represent saltation and bombardment processes. Larger soil particles, although not staying airborne, are essential because their impacts release smaller fragments that otherwise could not overcome interparticle cohesion forces via wind induced friction velocity only .

Consequently, a consistent application of the scheme is not possible with GMINER alone, since the database lacks information on the coarser soil fractions. Figure  illustrates this limitation for a representative Saharan grid cell, highlighting the substantial portion of the soil PSD that is missing when mineral fractions are restricted to only the two finest classes.

The emission scheme inherently favors silt-sized particles, which require relatively low wind stress for entrainment and are small enough to remain airborne, as illustrated by the threshold friction velocity function in Fig. 1 of. This aerodynamic preference becomes problematic when GMINER assigns minerals exclusively to silt or clay size classes. Treating these mineral-resolved PSDs as representative of the full soil PSD within the framework introduces systematic biases: clay-sized minerals experience artificially suppressed emissions due to high cohesion thresholds that rely on larger saltating particles to be broken, while silt-sized minerals are prone to overestimation in the absence of coarser, non-erodible grains. Accurate dust emission calculations therefore require a complete soil PSD, where larger particles mobilize finer fractions via saltation bombardment . This issue is compounded by the methodology used to construct GMINER: wet sieving mechanically disaggregates soil samples, increasing the apparent fraction of clay-sized particles relative to undisturbed soils . While airborne clay-sized fractions generally match source soils and are preserved during transport , silt-sized mineral distributions show poorer agreement, partly due to biases introduced by wet-sieving .

To address these issues, several modeling studies have incorporated mineral-specific transformations from soil to aerosol PSDs e.g.,. These approaches build on the Brittle Fragmentation Theory (BFT) of , which posits that for the finest particles, the emitted size distribution is largely independent of soil properties and wind speed. BFT conceptualizes emission as a sequence of energetic collisions between saltating aggregates, producing fragments that predominantly fall below a characteristic size threshold. The resulting particle number concentration N is inversely proportional to the square of the particle diameter Dp, modulated by an exponential cutoff controlled by a characteristic aggregate diameter Dc, as expressed in the following relation: 8dNdln⁡Dp∝1Dp2exp⁡-DpDc3forDp>Ds, where Ds represents a scalar diameter within the “indivisible” size range. Particles at or below this scale are considered primary grains resistant to further fragmentation due to inherent particle cohesion. Following , the probability of finding such indivisible grains at a specific size Ds is dictated by the PSD of the fully dispersed, wet-sieved soil. Incorporating this distribution, the emitted number size distribution is formalized as: 9dNdln⁡Dp=1cNDp2exp⁡-DpDc3∫0Dpp(Ds)dDs, where cN is a normalization factor, and p(Ds) represents the probability density function derived from the wet-sieved PSD. In this integral, Ds serves as the scalar integration variable representing the diameters of the constituent primary grains. This formulation captures the physical constraint that an emitted aggregate of size Dp can only be composed of indivisible components smaller than or equal to itself Ds≤Dp. The integral term (the cumulative distribution function of the soil minerals) thus reduces the emission of small particles if the soil lacks sufficient fine indivisible grains, while the exponential cutoff suppresses the emissions of very large diameters.

The BFT posits that the emitted number concentration of small dust particles is largely independent of the undispersed soil size distribution, with Dc estimated at 12±1 µm . In this framework, the PSD of emitted aggregates is skewed toward larger diameters compared to wet-sieved soil, consistent with observations by and , which indicate that saltation-driven fragmentation preserves a greater fraction of mass in the silt-size range than suggested by wet-sieved samples. further show that roughly 45 % of the silt-sized emitted mass originates from aggregated indivisible particles that would be classified as clay-sized in wet-sieved soils. To reconcile this discrepancy, they proposed an empirical correction that reallocates a fraction of clay-sized minerals from wet-sieved soils to the silt fraction, particularly phyllosilicates (illite, kaolinite, and smectite), which are categorized in GMINER exclusively in clay size classes but contribute to silt-sized emissions in reality.

This adjustment is quantified using Eq. () to derive a generalized ratio of emitted clay- to silt-sized particles, assumed to be spatially invariant and independent of local soil conditions. This simplifying assumption has been adopted in several modeling studies . However, it applies only to particles below 20 µm, as larger particles are more sensitive to site-specific emission factors such as wind speed, soil properties, and surface roughness. Since GMINER defines the silt class up to 50 µm, a further adjustment is needed for particles in the 20–50 µm range. addressed this by scaling the clay contribution using measurements from , which provided empirical data on the coarse-mode mineral distribution of freshly emitted dust. Using Eq. (), they estimated a clay-to-silt ratio of 0.05 for diameters below 20 µm. When integrated over the full 0–50 µm range and volume-normalized to account for the coarser fraction, this corresponds to approximately 1.3 % of the total emitted mass being attributed to the clay mineral fraction.

This methodology rests on two critical assumptions. First, it assumes that the PSD measured at Tinfou by is representative of other dust source regions. While this may not hold universally, the observed increase in silt fraction with particle diameter is consistent with the physical principle that the threshold friction velocity for emission decreases with particle size . The approach also assumes that the emitted PSD is primarily determined by mineral-specific fragmentation, neglecting wind speed variations. This assumption is reasonable for particles below 20 µm, where Eq. () offers a reliable approximation, but becomes less reliable for coarser particles where emission dynamics are more sensitive to environmental conditions. Second, the model neglects modifications to the PSD caused by gravitational settling during transport to the Tinfou observation site. This simplification is justified by focusing only on measurements taken during high-concentration dust events, which are presumed to reflect recently emitted aerosols with minimal transport-induced size sorting.

This methodology is implemented here as the “modified” approach, in which the dispersed soil size distribution from the GMINER SMA is adjusted to more realistically represent the aerosol size distribution observed in atmospheric dust. The formulation follows , based on BFT. From GMINER, the mass fraction of each mineral k in the conceptual clay (0–2.5 µm) and silt (2.5–50 µm) size categories is denoted by mkc(st) and mks(st), respectively. It should be noted that these two broad categories are used for the mineralogical reallocation process and are distinct from the five discrete aerosol transport bins defined in Table . These fractions are normalized for each soil type st (arid soil type provided by FAO74) such that: 10∑k=18mkc(st)=1and∑k=18mks(st)=1, with 8 minerals considered. To better reflect the emitted PSD, the BFT-based method distributes all minerals across both size categories, including those originally classified exclusively as clay- or silt-sized in GMINER.

To assign mineral fractions to the actual aerosol size categories, soil texture information from the SoilGrids database is incorporated. Let mc(sx) and ms(sx) denote the soil mass fractions of clay- and silt-sized particles, normalized such that: 11mc(sx)+ms(sx)=1.

The combined soil mineral mass fraction in each size category at each location is then computed by weighting the GMINER mineral fractions by the local soil texture proportions: 12mkc(st,sx)=mc(sx)mkc(st),mks(st,sx)=ms(sx)mks(st). This ensures that the total soil mineral mass fraction across both reallocation size classes sums to unity: 13∑k=18mkc(st,sx)+mks(st,sx)=1.

In this formulation, soil mineral mass fractions vary spatially according to two factors: the arid soil type (st), which governs mineralogical composition, and the soil texture class (sx), which determines the relative abundance of clay- and silt-sized particles.

Once the soil mineral fractions are obtained from Eq. (), the next step is to derive the emitted mass fraction of each mineral in the clay and silt size SMA categories. Denote by ϕc and ϕs the total emitted mass fractions of clay- and silt-sized aerosols, constrained by: 14ϕc+ϕs=1.

These totals are composed of contributions from all minerals k. Let ϕkc and ϕks represent the emitted clay- and silt-sized fractions attributable to mineral k. They satisfy: 15ϕc=∑k=18ϕkcandϕs=∑k=18ϕks, and consequently: 16∑k=18ϕkc+ϕks=1.

Thus, the emitted mineral mass fractions across all species and size classes sum to unity. Following the BFT-based formulation, the global clay-sized emission fraction is prescribed as ϕc=0.013, constant across all locations. The contribution of mineral k to clay-size emissions is therefore: 17ϕkc(st)=ϕcmkc(st),ϕc=0.013, meaning that the proportion of emitted clay-sized dust mirrors the clay mineral fractions of the fully dispersed soil. In this context, the fully dispersed soil represents a theoretical, maximally dispersed state, a conceptual limit where all aggregates are broken down into their primary, indivisible constituents.

With ϕc prescribed, the total silt-sized fraction follows implicitly as: 18ϕs=1-ϕc=0.987.

For each mineral, the emitted silt-sized fraction ϕks has two sources: (1) particles originally present in the soil in the silt-size range, and (2) primary clay minerals that are emitted within larger silt-sized aggregates during the saltation-driven fragmentation of the parent soil. To account for this, the model performs an empirical reallocation where a fraction of the soil's clay mineral mass is transferred into the silt-sized aerosol population. This is expressed as: 19ϕks=ϑγkmkc(st,sx)+mks(st,sx). where γk is a mineral-specific reaggregation coefficient quantifying the clay-to-silt transfer, and ϑ is a normalization constant ensuring that the sum of all ϕks equals ϕs.

For simplicity, γk is assumed constant across minerals, with the exception of feldspar, gypsum, and quartz. Feldspar is typically overrepresented in the silt fraction due to its fragmentation behavior, while gypsum is soluble and exhibits distinct disaggregation and spatial patterns . Quartz, which dominates the coarse end of the size spectrum and shows little evidence of disaggregation, is prescribed with γquartz=0.

Notably, feldspar and gypsum require special treatment since they are present in both clay and silt fractions of atmospheric aerosols e.g., but are absent from the clay fraction in the GMINER SMA. Consequently, Eq. () cannot be applied directly. Instead, their clay-size fractions are estimated following the approach of , which uses proxy minerals and scaling ratios: for feldspar, the extension into the clay fraction is scaled according to the quartz clay-to-silt ratio: 20mfeldsparc(st)=mfeldspars(st)⋅mquartzc(st)mquartzs(st), and for gypsum, the calcite clay-to-silt ratio is used: 21mgypsumc(st)=mgypsums(st)⋅mcalcitec(st)mcalcites(st). To conserve the total clay-size mass balance (Eq. ), the phyllosilicate fractions in the clay category are proportionally reduced by soil type.

Once these database gaps are addressed, the emitted silt-sized fraction of mineral k is expressed based on Eqs. () and () as: 22ϕks(st,sx)=ϑ(st,sx)γkmc(sx)mkc(st)+ms(sx)mks(st).

In this formulation, γk acts to suppress the silt-size emissions of minerals such as quartz, while extending the contribution of clay-rich species (e.g., phyllosilicates) into the silt range. This redistribution reflects the contrasting aggregation behaviors of different minerals: quartz remains largely intact during saltation, whereas clay-mineral constituents are emitted as part of larger silt-sized aggregates rather than as fully dispersed fine particles.

A schematic representation of the “modified” approach is shown in the lower panel of Fig. . In this framework, a substantial reduction of the quartz fraction is introduced, along with a redistribution of minerals across size classes. Specifically, phyllosilicates are incorporated into the silt fraction, while feldspar and gypsum are extended into the clay fraction. The application of Eqs. ()–() thereby entails a relative decrease of feldspar and gypsum within the silt-size range, resulting from the additional contribution of phyllosilicates.

After constructing the adjusted mineral soil PSD, which better represents the composition of mineral aerosols at larger sizes, it is incorporated into the COSMO5.05-MUSCAT emission scheme. The underlying emission algorithm remains unchanged from the implementation described by : the vertical emission flux in each MUSCAT size bin is scaled by the corresponding soil mineral fraction, with each mineral assigned to its own bin within the respective size category (see Table ). In the revised configuration, the mineralogical SMA is updated to reflect the modified PSD, and additional bins are introduced to represent minerals that now occur simultaneously in both clay and silt size ranges.

To verify the “modified” mineral approach, the scheme was compared against the MONARCH reference for Haplic Xerosols in the northwestern Sahara . Despite differences in bin resolution and emission parameterizations, both models demonstrate a consistent redistribution pattern, specifically, the shifting of phyllosilicates toward silt-sized categories and a proportional reduction in quartz. The resulting mineralogical shift from soil to emitted aerosol for this soil type is illustrated in Fig.  in Appendix B.

Mineralogical dust simulations were conducted using a consistent COSMO5.05-MUSCAT physical configuration. The experimental design varied across two primary dimensions: the simulation period and the mineralogical scheme. To match the observational campaigns, simulations were performed for January–February 2022 (DUSTRISK) and June–July 2022 (JATAC). For each of these periods, the two distinct mineralogical composition approaches, the “original” and “modified” schemes, were applied. The simulation experiments are summarized in Table .

The COSMO5.05-MUSCAT model domain is set up to cover the majority of the Sahara Desert and extend westward over the Atlantic Ocean to include the Cape Verde archipelago. The domain is bounded by 30.75° W–39.32° E and 38.49° N–0.38° S as represented in Fig. . Simulations are conducted at a horizontal resolution of 0.25° (approximately 28 km), with a vertical discretization of 40 layers. The lowest, surface level, model layer has a thickness of 20 m.

Meteorological initial and boundary conditions for COSMO5.05-MUSCAT are provided by the DWD in the form of 3 hourly meteorological fields. To maintain realistic atmospheric conditions, the model is re-initialized every 48 h using overlapping cycles. Each 48 h cycle begins with a 24 h spin-up phase, during which only the COSMO5.05 meteorological model is active. Following this spin-up, MUSCAT is coupled to COSMO5.05 for the remaining 24 h to simulate aerosol transport and interactions. Only the output from these second-day simulations, when both COSMO5.05 and MUSCAT are fully coupled, is used for analysis. Continuity between cycles is ensured by starting each new COSMO5.05 run 24 h prior to the end of the previous MUSCAT simulation, while MUSCAT continues using its own prognostic fields from the preceding cycle as initial conditions. The entire simulation period is preceded by a 15 d intial spin-up, and results are outputted at 1 h intervals.

The MUSCAT dust emission scheme is controlled by external soil-related datasets. Vegetation cover is prescribed using the FCOVER product from the Copernicus Global Land Service , which provides fractional green vegetation coverage. Soil moisture input is taken from ERA5-Land hourly reanalysis data , specifically utilizing the volumetric soil water content of the 7 cm uppermost soil layer. Soil texture and mineralogical composition are provided by the SoilGrids and GMINER databases , respectively, and implemented via the two distinct modeling approaches detailed in Sect. . Aerodynamic roughness lengths are provided through the dataset of .

These datasets, combined with the model's spatial constraints, restrict dust activation exclusively to continental source regions. Furthermore, the spatial distribution of active dust emission sources is constrained by the MSG-SEVIRI dust source activation frequency map by .

During the JATAC 2022 campaign, in-situ recollection of mineral dust aerosols were performed using Unmanned Aerial Vehicles (UAVs) over São Vicente, Cape Verde, (indicated by the yellow star in Fig. ) throughout June 2022. The Cyprus Institute conducted 25 UAV flights, each equipped with Optical Particle Counters (OPCs) for fine- and coarse-mode height-resolved PSD observations. Dust particles were collected from the atmospheric layers using onboard impactor samples , with a Giant Particle Collector (GPaC) capable of capturing particle diameters from nanometers up to tens of micrometers . The GPaC's upper size limit depends on airspeed, pressure, and temperature; as a reference, during the SAMUM-2 campaign, particles up to 28.5 µm were successfully collected .

Two OPCs were mounted on each UAV. The Universal Cloud and Aerosol Sounding System counted aerosols with diameters from 0.28–17.0 µm, while the Printed Optical Particle Spectrometer measured number concentrations in the 0.14–3 µm range . OPC measurements were used to validate PSDs derived from collected dust samples, following the methodology of .

Aerosol chemical composition and single-particle characteristics were analyzed using a scanning electron microscope coupled with Scanning Electron Microscope-Energy Dispersive X-ray spectrometry (SEM-EDX) to determine particle size, shape, and elemental composition. Back-scattered images were used to study particles with projected area diameters (PAD) >0.5µm, where PAD (Dp=4Ap/π, with Ap as the particle area) closely approximates aerodynamic diameter for dust . SEM-EDX provides normalized atomic percentages for elements including: F, Na, Mg, Al, Si, P, S, Cl, K, Ca, Ti, V, Cr, Mn, Fe, Zn, and Pb. This method allows detection of particles up to 30 µm, with measurement precision for major compounds within 2 % relative standard deviation (RSD), while minor compounds range from 10 %–20 % for particles>3µm, and can exceed to 100 % for the smallest particles. Diameter measurement uncertainty decreases with size, from ∼1.5% RSD at 2 µm to <1% for particle diameters>3µm .

Impactor sampler collected 18 samples, and analyzed size, shape and elemental composition for 25 000 single particles during the measurement period. Samples with a low number of particles (<500) were discarded to ensure reliable statistics while counting/grouping each particle group. This screening protocol resulted in a final validation dataset comprising 17 high-quality samples.

To enable a consistent comparison between measurements and model output, the observed minerals are classified into the same mineral classes available in the GMINER soil database. This ensures that the measured and simulated mineral categories are directly comparable, avoiding ambiguities arising from differences in mineral classification. Size-segregated chemical compositions (same as Table ) were determined for the primary dust aerosol minerals: illite, kaolinite, smectite, quartz, feldspar, calcite, gypsum and hematite were derived by defining ideal elemental compositions for each modeled mineral, accounting for common impurities associated with each mineral phase (Table  in Appendix A). Individual measured particles are then matched to the nearest ideal composition, and counting statistics for each mineral group are used to estimate uncertainty via two-sided 95 % confidence intervals under a binomial assumption.

For iron oxide minerals such as hematite, the ratio reported in Table  reflects the iron oxide content from a single-particle perspective. However, a significant fraction of iron in dust occurs within the crystal lattice of other minerals , necessitating an alternative approach to estimate total iron oxide content.

Following , we use two parameters: (i) the total iron oxide percentage (Feoxides), which accounts for iron present either as pure oxy-hydroxides (e.g., hematite, goethite) or incorporated within mineral lattices, and (ii) the total Fe index, defined as the atomic ratio of Fe to the sum of all quantified elements. The total iron oxide percentage is calculated as: 23Feoxides=mFeox%mFe%⋅MFeMdry⋅100, where mFeox% and mFe% are the mass fractions of iron oxides and elemental iron relative to the total dust mass, respectively, obtained from Table 3 in for different Saharan source regions. MFe is the estimated Fe mass within a single particle from SEM-EDX measurements, and Mdry is the particle's dry mass.

Using this formulation, mass percentages for both total iron oxide and hematite are derived. For hematite specifically, hematite mass fraction from is used to substitute mFeox% in Eq. (). The uncertainty arising from the representative ratio of iron oxide content from 19 desert dust types given by Table 3 in which is generally less than 15 %. For us, this is a systematic error, as we apply the same ratio to all samples.

Additionally, elemental mass concentrations are provided for the species most commonly associated with mineral dust, including Si, Al, Fe, Ca, K, Mg and S. Both mineralogical and elemental data are partitioned into the MUSCAT size bins up to BIN26.

The DUSTRISK 2022 campaign took place in the Cape Verde archipelago during January and February 2022. Particulate matter was sampled at two distinct sites, referred to as inflow and outflow, on the island of Santiago (Cabo Verde). The inflow site (14°54^′32^′′ N, 23°30^′54^′′ W, indicated by the violet × in Fig. ) was positioned in the city of Praia, on the northeastern coast, directly facing the ocean.

The sampling site was strategically positioned to intercept air masses dominated by the prevailing northeast trade winds, which transport Saharan mineral dust directly from the African continent. Located upwind of both the urban center of Praia and the island's interior, the site is well-situated to ensure that collected aerosols represent long-range continental transport with minimal interference from local island soils. Although minor anthropogenic contributions from nearby urban areas cannot be entirely discounted, the site's orientation ensures that local influences remain negligible, as corroborated by previous studies . In contrast, the outflow site is positioned downwind of the major metropolitan area and is significantly more susceptible to local anthropogenic and local dust emissions . To ensure direct comparability with model simulations that represent only continental dust sources, this analysis focuses exclusively on measurements from the inflow site.

Particulate matter (PM) samples were collected using Digitel DHA-80 high-volume samplers (Walter Riemer Mess Technik, Germany) operating at an average flow rate of 500 Lmin-1. At each site, separate PM2.5 and PM10 samplers were deployed. Sampling typically occurred over 24 h periods (noon to noon); however, during pronounced dust events, sampling duration was shortened to capture peak concentrations. All samples were collected on Ahlstrom micro-quartz fiber filters (MK 360) .

Elemental concentrations were determined using Total reflection X-Ray Fluorescence (TXRF) spectroscopy, following the methodology described by . Filter samples of 1.5 cm2 area were prepared by digesting three 8 mm spots per filter in a HCl/HNO3 mixture (0.375 and 1.125 mL, respectively) using a microwave digester (Mars 6, CEM, Germany). Internal standards (Sc/Co at 10 µL) were added to the digested solutions, which were then applied to siliconized quartz carriers, dried at 80 °C, and analyzed using a Bruker S4 T-STAR instrument. This procedure determined concentrations of Al, Ti, Ca, K, Mg, S, Si, Cr, Mn, and Fe. To ensure precision and account for potential sample heterogeneity, each carrier was analyzed at two rotational positions (0 and 90°). Measurement uncertainty was quantified by calculating the relative error between these replicates, revealing a dependency on mass loading: uncertainties reached up to 15 % for samples with very high dust load (>150µgm-3) while remaining below 6 % for samples with low dust concentrations typically below 38 µgm-3. These variations are attributed to instrumental limitations previously characterized in .

Notably, blank silicon measurements for the filters and TXRF sample carriers showed significant, non-reproducible variations due to inconsistent filter densities and surface reflectance. The resulting uncertainties were too high for a meaningful comparison; therefore, silicon was excluded from the analysis. To isolate the mineral dust signal from the background aerosol, the total mineral dust mass concentration was reconstructed from the remaining elemental data following the procedure of . In this approach, the concentrations of key crustal elements (Al, Fe, Ca, Ti) are adjusted for their common oxide forms to estimate the total mineral mass. The dust concentrations were quantified using the stoichiometric relationship for crustal elements following and : 24MD=1.16×((1.90×Al)+(23.3×Ca)+(2.09×Fe)+(1.67×Ti)).

Following established classification for mineral dust events in the Cabo Verde region , a mineral dust mass concentration threshold of ≧38µgm-3 was applied. The threshold was chosen based on long-term analysis of PM10 concentrations at the CVAO sampling site, during which dust concentrations were consistently high. This threshold, combined with air mass back-trajectory analysis, serves to distinguish significant long-range transported dust events from maritime background or local aerosols. This screening resulted in a final validation set consisting of 34 PM2.5 and 45 PM10 filter samples.

To ensure a robust comparison between COSMO5.05-MUSCAT simulations and in-situ observations, the model output is spatio-temporally sampled to align with the specific characteristics of each measurement campaign. Spatially, model variables were extracted from the grid cells corresponding to the geographic coordinates of the sampling sites. Temporally, hourly model outputs were averaged over the specific integration times of the filter samples. This alignment accounts for varying sampling durations, ranging from the high-frequency UAV cycles during JATAC 2022 to the day-long ground-based periods of the DUSTRISK 2022 campaign. To ensure temporal concurrency, a rounding convention is applied: if a sampling interval began or ended within 30 min of a full hour, the nearest hourly model timestep is used to define the averaging window. For example, a sampling period from 10:00–12:15 GMT (time zone throughout) is matched with the model average from 10:00–12:00, while a period from 10:10–12:36 is matched with the 10:00–13:00 simulated average.

For the compilation of North African mineralogical measurements, where exact temporal matching is often not possible due to the lack of concurrent simulated data, the time dimension is constrained by selecting observations from the same seasons as the simulation period. Furthermore, the comparison is restricted to sites and dust trajectories similar to those simulated, ensuring that the emission sites to measurement place pathways in the model are representative of the taken measurements, as detailed in the following section.

It is important to note that the “original” mineralogy scheme possesses structural constraints regarding the simulation-to-measurements comparisons. Specifically, phyllosilicates are restricted to the clay and fine-silt fractions (BIN01 and 1/2 BIN03), while minerals such as feldspar and gypsum are defined with no mass in the clay-sized fraction (BIN01). These inherent model configurations limit the scope of certain comparisons where observations report these minerals in size fractions not supported by the scheme's internal mapping.

Beyond these structural constraints, the definition of the simulated mineral mass concentrations also requires clarification. The simulated mineral mass fraction is defined as the ratio of a given mineral's mass to the total dust mass fraction within the size range used for comparison. For example, for bulk comparisons, this ratio is computed by summing the mineral mass across all size bins in which it is simulated and dividing by the total dust mass summed across all size bins. Importantly, this definition is applied consistently even when a mineral is not represented across the full size spectrum; in such cases, the denominator still comprises the total dust mass over all size bins. This approach implicitly assumes that the simulated minerals collectively account for the entirety of the dust mass. However, it should be noted that additional minerals present in the observations but not represented in the model also contribute to the measured total dust mass. Since these non-simulated minerals cannot be quantified within the modeling framework, their contribution introduces and inherent but unquantifiable offset between the simulated and observed mineral mass fractions.

The agreement between simulated (y) and observed (x) mineralogical and elemental concentrations is quantified using the Mean Bias (MB), the Pearson correlation coefficient (r), and an orthogonal distance regression. The correlation coefficient r serves as a measure of spatial agreement, evaluating the model's ability to replicate the geographical variability and regional distribution of mineral dust signatures across the study domain. The Pearson correlation coefficient is defined as: 25r=∑i=1n(yi-y‾)(xi-x‾)∑i=1n(yi-y‾)2∑i=1n(xi-x‾)2, where y‾ and x‾ represent the mean values of the simulated and observed datasets, respectively, and n is the number of paired spatial samples.

To account for uncertainties inherent to the baseline measurements (x), the linear model is fitted using an orthogonal distance regression rather than ordinary least squares, which minimizes residuals in both axes simultaneously and thereby prevents attenuation bias in the slope estimation. The linear model takes the form: 26y=mx+b, where m is the slope and b is the intercept, obtained by minimizing the weighted sum of squared orthogonal residuals: 27min⁡m,b∑i=1nwx,ixi-yi-bm2, where wx,i=1/σx,i2 are weights inversely proportional to the observational variance σx,i2 of each measurement.

The slope and intercept derived from this fit are reported alongside the Mean Bias (MB) in the results tables. The MB identifies systematic over- or underestimations and is defined as: 28MB=1n∑i=1n(yi-xi).

Model outputs from DUST-CM-01 and DUST-CM-02 (see Table ) are compared against a compilation of in-situ aerosol mineralogical measurements. The simulated period (January–February 2022) was specifically selected for two reasons. First, it enables a direct comparison with the “original” mineralogy scheme presented by , ensuring that any differences in model performance can be attributed to the updated mineralogical parametrization rather than to variations in meteorological forcing. Second, although an additional modeling period covering Northern Hemispheric (NH) summer conditions is available, it was not employed here since the measurement campaigns most representative of NH summer dust transport were already matched to source regions and transport pathways compatible with the January–February period through our classification methodology, rendering the additional period unnecessary for the present evaluation.

The core of this compilation is based on the global dataset of , subsetted to include only locations within the modeling domain, and further expanded with one additional dataset from the Atlas Mountains region in Morocco . Measurements predating the 1980 s were excluded, as the absence of satellite imagery prior to this period prevents reliable identification of dust emission events and source regions. This expanded dataset is publicly available via the zenodo repository (10.5281/zenodo.21132304, ), and the corresponding measurement locations are represented by blue dots in Fig. .

The subsequent systematic processing is illustrated in Fig. , which delineates the pathways for both in-situ observations (gray arrows) and COSMO5.05-MUSCAT simulation results (green arrows). The first decision step depends on the number of distinct dust events captured by each dataset. Measuring campaigns that sampled more than two distinct dust events are categorized as long-term campaigns, for which average mineral mass fractions at the measurement location grid cell are computed over the entire simulation period and compared directly against observations. This approach was applied to the datasets of , , , , and .

In contrast, shorter campaigns covering two or fewer dust events required a more detailed matching procedure, initiating the “dust event analysis” workflow illustrated in Fig. . This methodology integrates HYSPLIT back trajectories with satellite-derived dust emission records from the MSG-SEVIRI Dust RGB product (EUMETSAT) to identify probable source regions, following the approach of , recently applied in .

To ensure a like-for-like comparison, simulated dust plumes were selected only when their transport pathways and source regions were consistent with the specific observed events, with model output extracted from the grid cell corresponding to the geographic coordinates of each measurement site. The “dust event analysis” workflow, illustrated in Fig. , begins by identifying the probable source region of the observed sample using HYSPLIT back-trajectories intersected with dust emission records from the MSG-SEVIRI Dust RGB product. Simultaneously, backward trajectories for simulated dust events were computed using the LAGRANTO tool , driven by COSMO5.05 meteorological fields. Comparison with measurements proceeded only when the simulated and observed plumes originated from the same dust source region, thereby allowing for the direct matching of individual plume signatures with high-resolution model outputs.

Three campaigns required this rigorous event-based matching procedure: , , and . For , which predates the MSG-SEVIRI record, Meteosat-7 natural color imagery was utilized to track the specific plume transported from southern Algeria to Tenerife on 29 July 2002. For the remaining campaigns, observed plumes were individually matched to simulated events by ensuring that both the HYSPLIT (observational) and LAGRANTO (model) back-trajectories intersected with the same active MSG-SEVIRI dust emission areas. The resulting compilation spans multiple seasons and years; while Saharan dust transport pathways exhibit pronounced seasonal variability , the event-based matching procedure applied to the NH summer measurements ensures that the simulated and observed plumes originate from the same source region, mitigating potential biases introduced by seasonal differences in dust emission patterns.

To ensure a robust comparison between varying measurement techniques and model outputs, size-matching was performed based on the reported diameter ranges. For each comparison, the simulated mineral mass fraction was calculated by integrating the mass of a specific mineral across the relevant model bins and dividing this sum by the total simulated dust mass within those same bins. Minerals absent from a given size bin in the model are excluded from the comparison at that size, meaning that differences in mineral implementation between schemes affect the number of available data points. For campaigns reporting bulk measurements, model values were integrated across all available size bins (BIN01 through BIN80). For measurements classified as “clay”, the mass fraction was derived from BIN01+1/2BIN03, while “silt” classifications were compared to the average of 1/2BIN03+BIN09+BIN26+BIN80. For the campaign reporting high-resolution size distributions , model bins were matched to the closest reported diameter ranges: 0–1 µm is compared to BIN01; 1–2 µm to BIN03; 4–8 µm to BIN09; 8–16 µm to BIN26; and 32–64 µm to BIN80.

For the evaluation metrics, measurement sites are grouped by geographical region to reflect the commonality of dust sources and transport pathways. Group 1 comprises sites in the Sahel, specifically Niger and Nigeria; group 2 covers the receptor islands of the Canary Islands and Cabo Verde, and group 3 gathers sites surrounding the Atlas Mountains in Morocco. Given the limited number of data points within each group, the linear fit and correlation coefficient do not provide statistically meaningful information, and only the mean bias is reported for these grouped comparisons.

Following these size-matching and spatio-temporal constraints, the final size selected mineral mass fraction comparison is performed using the model output from the specific grid cell corresponding to each measurement site. While measurement uncertainties were not consistently reported across all original studies, they have been included in the comparison with simulated data whenever available.

Despite these rigorous criteria, inherent uncertainties remain. Meteorological anomalies or atypical seasonal transport patterns in the historical data may not be fully captured by the January–February 2022 simulation period. Furthermore, several sites are located near major dust sources where local emissions may dominate the observed mineralogy in ways the model's spatial resolution cannot fully resolve e.g.,. Finally, inconsistencies in particle size classification across datasets, such as varying definitions for “bulk”, “clay”, and “silt” fractions, introduce additional uncertainty. While model bins were mapped to these categories as previously described, the sensitivity of these boundaries (e.g., 2.0 vs. 2.5 µm) highlights the necessity for high-resolution, well-characterized, and concurrent datasets, such as those obtained during the DUSTRISK 2022 and JATAC 2022 campaigns.

To evaluate the performance of the “modified” mineralogical scheme, we compare the simulated dust composition against two distinct observational benchmarks. First, we assess the model's ability to capture the general characteristics of North African mineral dust using a comprehensive compilation of historical aerosol measurements from across the region as described in Sect. . This provides a baseline for regional performance. Second, we evaluate the scheme's precision during a specific transport event using high-resolution mineralogical data from the JATAC 2022 campaign. Together, these comparisons provide a complementary validation of the model's capacity to represent both the broad regional characteristics and the episodic variability of dust mineralogy.

Simulated mineral compositions are converted into elemental abundances via the relationships in Table , providing a complementary and independent metric for evaluating model performance. While size-resolved elemental comparisons broaden the assessment, they introduce additional uncertainties: despite filtering for dust-dominated samples, contributions from minerals outside the modeling framework cannot be fully excluded. This is particularly relevant for the DUSTRISK 2022 dataset, collected near the surface during NH winter, where interference from local emissions or Sahelian biomass burning plumes is more likely ; the JATAC 2022 observations, made within the Saharan Air Layer during NH summer, are more representative of pure Saharan dust, though minor contamination during transport cannot be ruled out. It should also be noted that the simulated elemental compositions are sensitive to the assumed elemental fractions per mineral, for which different choices have been made across modeling studies ; the compositions adopted here were selected as most representative of the mineralogical signatures expected at the Cabo Verde receptor site, as discussed in Sect. .