The Weather Research and Forecasting model coupled with Chemistry (WRF-Chem v4.6.0) was employed to simulate a set of dust events over the RM using the Goddard Chemistry Aerosol Radiation and Transport (GOCART) aerosol scheme, which explicitly represents the emission, transport, deposition, and removal of dust particles .

The model configuration consists of three domains with horizontal resolutions of 9 km (d01, North Africa and central Europe), 5 km (d02, IP), and 1 km (d03, RM), covering North Africa and southern Europe, the IP and the RM, respectively (Fig. ). The model employs 30 vertical levels extending up to 10 hPa, with enhanced resolution near the surface to better resolve boundary layer processes critical for dust emission and transport.

The physical parameterization includes the Morrison two-moment microphysics scheme, the Grell–3D cumulus parameterization (applied to domains d01 and d02 only, as convection is explicitly resolved in the 1 km domain), the RRTMG shortwave and longwave radiation schemes, the Yonsei University (YSU) planetary boundary layer scheme, and the Noah land surface model. A complete list of model options and references is provided in Table .

The chemical mechanism selected was the simplified GOCART scheme without ozone chemistry (WRF option 300). The data used to provide initial and boundary meteorological conditions (IC and BC, respectively) for the outer domain (d01) were obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) fifth-generation reanalysis dataset (ERA5), which provides global coverage at a horizontal resolution of 0.25° and an hourly temporal resolution .

The performance of the newly developed erodibility datasets was evaluated over several simulation periods (Table ). The simulation periods were not selected to provide a continuous time series but as a set of well-defined case studies representative of the main mechanisms that drive PM₁₀ episodes in the RM. Specifically, the five periods were chosen to sample above-average PM₁₀ events covering the two dominant regimes in the region: long-range Saharan dust advection (March 2022, November 2025) and weak-forcing situations favoring local dust uplift and accumulation (July 2022, August 2025). Additionally, an intermediate situation combining convective dust transport with local uplift processes was included (April 2025). The selection was further constrained by the simultaneous availability of quality-controlled observations across the monitoring network, which is why no episode from 2023–2024 is included. Regarding representativeness, although the five episodes span different seasons and synoptic regimes and are evaluated against eight stations, they constitute an event-based sample. Therefore, the study should be understood as applying to high-PM₁₀ dust episodes rather than to the full climatological PM₁₀ distribution.

To provide meteorological context and characterize the different dust episodes, a synoptic-scale analysis was conducted using daily mean fields of sea-level pressure (SLP), geopotential height at 500 hPa (Z500), and temperature at 850 hPa (T850) derived from the ERA5 reanalysis dataset. Figure  presents the daily mean synoptic fields for the peak PM₁₀ day of each episode together with the surrounding days, illustrating the large-scale atmospheric conditions associated with the dust events. Additionally, Figs.  and  show the corresponding daily mean PM₁₀ distributions simulated using the SOILHD and CTRL datasets, respectively. The synoptic analysis indicates that three of the analyzed episodes (March 2022, April 2025, and November 2025) are associated with large-scale, low-pressure systems over the Atlantic Ocean that promote the transport of Saharan dust toward the IP. In particular, during the March 2022 event, the low-pressure system was centered over northern Morocco, enhancing dust advection into the study region.

In contrast, the remaining two episodes (July 2022 and August 2025) occurred under relatively stable atmospheric conditions characterized by weak synoptic forcing and low wind speeds, favoring the accumulation of PM₁₀. For the August 2025 episode, the simulations reveal a pronounced local dust emission event over the southeastern IP, as shown in the PM₁₀ distribution maps. This behavior is consistent with the observational records, which indicate maximum PM₁₀ concentrations of up to 204.5 µg m⁻³ during the episode. This event is therefore interpreted as a local dust uplift episode and was selected for a more detailed analysis. In particular, it provides a suitable framework for assessing the capability of the newly developed SOILHD dataset to represent local-scale variability in dust emissions and its subsequent impact on PM₁₀ concentrations. For this reason, a substantial portion of the results discussed in this study focuses on the August 2025 simulations, although all analyzed periods are evaluated throughout the manuscript. For each simulation, the first day was treated as the model spin-up and excluded from the analysis to ensure numerical stability and reduce initialization effects.

For each period, three experiments were conducted. The first, denoted as the control (CTRL) simulation, employed the default WRF-Chem erodibility dataset. The second, referred to as EROD-HR, utilized the newly developed high-resolution EROD dataset. Finally, the SOILHD experiment incorporated the further enhanced Soil-Type Heterogeneous Dataset. All simulations were run with identical physical and chemical configurations (Table ), differing only in the erodibility dataset employed to isolate its effect on dust emissions and transport. For the August 2025 period, an additional simulation was performed, featuring the soil erodibility dataset developed by , to account for its performance in our domain of simulation. We simply refer to this simulation as LI2023, and its results can be observed in Sect.  and  of this work. Dust emissions were computed online using the GOCART scheme, based on 10 m wind speed, soil moisture from the Noah land surface model, and the erodibility factor from the respective dataset, according to Eq. ().

In the GOCART scheme, for a given size bin p, the dust emission flux value (kg m⁻² s⁻¹) is obtained for each surface model cell using the following equation : 1Fp=CSspU2U-Ut(Dp,θs),U>Ut(Dp,θs) where C is a dimensional proportionality constant, S is a unitless dust source strength function, sp is the mass fraction of emittable dust from a given dust bin at the surface, U is the horizontal wind speed at 10 m, and Ut is the threshold 10 m wind speed required to initiate erosion for a given particle diameter Dp and soil surface wetness θs. Fp is 0 in any other case where U does not exceed the threshold velocity. Once Ut is surpassed, the dust source function S arises as the key parameter to determine the location and amount of dust emitted. S was originally computed by based on the idea that dust material is often generated in alluvial processes and accumulates in low points. For this reason, S was determined based on the degree of topographic relief surrounding a model grid cell using the following equation : 2S=zmax-zizmax-zmin5 where zi is the elevation of the cell, and zmax and zmin are the maximum and minimum elevation in a 10°×10° window centered on the same grid cell. S is set to 0 anywhere bare soil is not indicated by the Advanced Very High Resolution Radiometer (AVHRR) data . At the time of this study, the GOCART scheme in WRF uses a precalculated source strength function S, stored in the code as the variable EROD, which is read in and interpolated to the model grid by the WRF-Chem preprocessor . This static field was calculated by applying Eq. () to a 0.25° resolution elevation dataset. This results in a coarse-resolution dust source dataset, which limits its ability to represent local-scale dust sources accurately (Figs.  and ).

In a first step to overcome the spatial limitations of the default 0.25° erodibility field, a new high-resolution dataset (hereafter EROD-HR) was developed following the original formulation of the dust source strength function (Eq. ), while incorporating higher-resolution and more physically consistent geophysical inputs. The main objective was to improve the spatial representativeness of potential dust sources, particularly in regions where local terrain variability, land cover, and soil texture exert a dominant control on dust uplift. The dust source strength function S was recalculated globally using the GMTED2010 elevation model . The native 7.5 arcsec (≈ 250 m) topographic data were resampled to a 0.0625° (≈ 5 km) spatial resolution to balance physical detail and computational feasibility for regional WRF-Chem applications. For the inner domains covering the IP and the RM, a 1 km version was generated to enable fine-scale validation and to assess the impact of convective-scale topographic variability on dust emission.

To ensure physical consistency with the GOCART emission formulation, only grid cells identified as barren or sparsely vegetated were considered active dust sources. Land cover classification was derived from the MODIS 20-category global land use dataset , and non-erodible classes (e.g., urban areas, forests, croplands, permanent water bodies, and snow or ice) were masked by setting S=0. This ensures that dust uplift is restricted to areas where bare-soil exposure is physically plausible. The resulting dataset preserves the conceptual structure of the original erodibility field while substantially improving its spatial fidelity.

In WRF-Chem, soil texture is represented in a simplified manner within the dust emission formulation. The erodibility field is defined by three layers corresponding to the soil textural classes contributing to dust emission: sand, silt, and clay. In the standard scheme , constant global mass fractions are assumed for all grid cells (50 % sand, 25 % silt, and 25 % clay). Accordingly, the EROD-HR dataset was constructed by multiplying the recalculated S field by these fixed fractions. As a result, EROD-HR remains fully compatible with the internal WRF-Chem structure while incorporating higher-resolution topographic and land-cover information. This 50–25–25 % sand-silt-clay assumption has also been adopted in other recent erodibility datasets based on satellite-derived indicators of soil erosion and vegetation scarcity . Nevertheless, assuming uniform soil composition can lead to unrealistic dust emission strengths in regions with pronounced soil heterogeneity. To address this limitation, an enhanced erodibility dataset, named SOILHD (Soil-Type Heterogeneous Dataset), was developed by replacing the fixed soil fractions with spatially variable, observation-based soil composition fields (Fig. ).

SOILHD data were derived from two complementary sources. First, the global map of Local Climate Zones (LCZs) developed by was used to identify potential dust source areas while excluding surfaces erroneously classified as bare soil in previous datasets, such as high-mountain regions, urban areas, snow- and ice-covered surfaces, and greenhouses. This dataset consists of a global LCZ map at 100 m spatial resolution, derived from multiple Earth observation products and LCZ class labels . The LCZ typology distinguishes urban and natural surface types based on characteristic combinations of land cover and physical properties. Of the 17 LCZ classes, 10 correspond to built environments, while the remaining seven represent natural land-cover classes. For this study, the natural barren-soil class was extracted and conservatively remapped to a 1 km resolution grid, yielding a global map of barren-soil fraction for each grid cell.

Secondly, the ISRIC SoilGrids250m database , which provides global estimates of soil properties (including sand, silt, and clay fractions) based on machine-learning models trained on more than 230 000 soil profile observations at 250 m spatial resolution and multiple depth layers (0–200 cm), was employed. Only the 0–5 cm layer was used in this study. The effective radii associated with the silt and clay fractions correspond to the GOCART dust size bins: bin 1 for clay (0–1 µm) and bins 2–4 for silt (1–10 µm). For each soil texture class c (sand, silt, clay), a spatially explicit mass fraction field fc(x,y) was computed by averaging the 0–5 cm layer and converted into binary format. The erodibility layer for the soil class c for the SOILHD dataset (Ec) was then defined as: 3Ec(x,y)=fc(x,y)⋅B(x,y) where B(x,y) denotes the barren-soil fraction at each grid point. For the SOILHD dataset, Ec is equivalent to the source strength function S in Eq. (). This formulation introduces realistic spatial variability in soil composition while restricting dust emission to locations where surface conditions are suitable for dust uplift at a 1 km scale. The resulting dataset provides three distinct erodibility layers per grid cell with physically based weighting factors that reflect local soil texture, rather than relying on globally uniform constants. This refinement enhances the ability of WRF-Chem to represent regional heterogeneities in dust emission intensity, particularly across transitions between sand-dominated and fine-grained soils.

Both the EROD-HR and SOILHD datasets were reprojected onto a regular latitude-longitude grid, converted to binary format, and stored following the same three-dimensional structure as the default erodibility file to ensure direct compatibility with the WRF Preprocessing System (WPS).

Finally, the new datasets were validated against the original erodibility field to ensure structural continuity (see Sect. ), and the SOILHD dataset was rescaled to match the order of magnitude of the default erodibility values (see Sec. ). Test simulations confirmed that both datasets were correctly ingested and interpolated by WRF-Chem without requiring any modifications to the dust emission code. The EROD-HR dataset serves as a baseline high-resolution product emphasizing improved topographic and land-cover representation, whereas SOILHD further incorporates soil-texture heterogeneity to refine the physical characterization of dust source potential.

Figures and show an RGB representation of the different datasets for domains d01 and d03, respectively. The SOILHD dataset introduces a clear modification in the qualitative soil texture distribution compared to both CTRL and EROD-HR configurations. The inclusion of soil-type heterogeneity leads to a more realistic spatial redistribution of erodible fractions. In domain d01, all datasets preserve the dominant sandy character of the Sahara region; however, SOILHD introduces a more structured and physically consistent presence of silt- and clay-enriched areas compared to the more uniform CTRL and EROD-HR configurations. In contrast, EROD-HR mainly preserves the spatial pattern of CTRL while increasing resolution, without substantially altering the underlying soil composition. The SOILHD dataset, instead, redistributes the erodibility in a way that better captures localized heterogeneity associated with fine-particle availability.

The differences become more pronounced in domain d03 (RM), where SOILHD exhibits a clear predominance of silt-rich soils, in contrast to the 50–25–25 % assumption embedded in the CTRL and EROD-HR configurations. This indicates that fine particles play a more relevant role in this region than originally assumed, leading to a more heterogeneous and spatially structured erodibility field. In comparison, LI2023 shows a markedly different behavior, with an almost uniform and higher erodibility across large portions of the Sahara under the default 50–25–25 % assumption, but also with unrealistic values over mountainous regions such as the Alps and the Pyrenees, where soil conditions should not support dust emissions. Conversely, in domain d03, LI2023 exhibits very limited active source areas, resulting in a nearly suppressed local dust emission potential, which is not consistent with the semi-arid and sparsely vegetated characteristics of the RM.

The tuning of soil erodibility is a common practice in dust modeling, particularly when introducing new erodibility datasets or transferring models across regions with contrasting soil, land-use, and climatic characteristics. Previous studies (e.g. ) have shown that moderate, physically motivated adjustments to erodibility can substantially improve the representation of dust emissions and downwind concentrations.

To allow a simple, systematic tuning of the erodibility field, we developed the EROD Tuner tool . This utility operates directly on the EROD variable contained in the WPS geo_em.d0n.nc files, modifying the soil erodibility prior to the WRF-Chem integration. Working at the preprocessing stage ensures that the modified erodibility is consistently propagated through the entire modeling workflow, while allowing different tuning strategies to be applied independently to each model domain.

A key feature of EROD Tuner is its flexibility. The tool supports a wide range of conversion functions, from simple linear scaling to nonlinear formulations targeting specific portions of the erodibility distribution. These include polynomial functions that damp extreme values, power-law transformations that enhance weakly erodible soils, logarithmic and sigmoid functions introducing saturation effects, and hybrid formulations that smoothly combine linear and nonlinear behavior. This flexibility allows the erodibility response to be tailored to different physical hypotheses, such as increased soil cohesion, threshold-like transitions between erodible and non-erodible surfaces, or the suppression of unrealistically strong source areas.

All conversion functions are formulated in a generic, dimensionless form, making them applicable to erodibility datasets beyond those presented here. A dedicated namelist allows the selection of different functions and parameters for each nested domain, enabling region-specific tuning while preserving a unified modeling framework. This design facilitates sensitivity analyses, uncertainty assessments, and intercomparison studies in which the impact of erodibility assumptions must be isolated. The code also allows straightforward implementation of additional custom functions. The EROD Tuner tool is fully open-source and is publicly available on Zenodo (10.5281/zenodo.18606994, ).

In the present study, the tuning applied to the SOILHD erodibility dataset was intentionally conservative. The adjustment was limited to a uniform linear scaling of the total erodibility, defined as the sum of the sand, silt, and clay components. A constant multiplicative factor of 10⁻¹ was applied to bring SOILHD erodibility values into the same order of magnitude as the default WRF erodibility field. This approach preserves the spatial variability and relative contrasts introduced by SOILHD while ensuring comparability with the standard model configuration.

Modeled dust concentrations (PM₁₀) were evaluated against ground-based observations within the innermost model domain (d03) using data from the Regional Air Quality Monitoring Network of the RM (SINQLAIR) . SINQLAIR is operated by the regional government and provides high-quality, real-time measurements of major atmospheric pollutants for regulatory assessment and public information, with a native temporal resolution of 10 min.

The network currently consists of 11 fixed monitoring stations distributed across urban, industrial, and rural environments, as well as one mobile unit deployed during specific campaigns or pollution episodes. SINQLAIR routinely measures concentrations of particulate matter (PM₁₀ and PM_2.5), nitrogen oxides (NO_x, NO₂), sulfur dioxide (SO₂), tropospheric ozone (O₃), carbon monoxide (CO), and other trace pollutants.

Based on data availability for the simulated periods, 8 stations were selected for the present analysis. These stations are listed in Table , and their geographical locations are shown in Fig. . To ensure temporal consistency with the model output, the observational time series were aggregated to hourly mean values.

To quantify the model performance and the impact of using the high-resolution EROD-HR and SOILHD datasets, the following statistical metrics were calculated. On one hand, the Pearson correlation coefficient (R) is used to assess the temporal agreement between observed and simulated concentrations. On the other hand, Mean Bias Error (MBE) and Root Mean Square Error (RMSE) are used to evaluate the accuracy and systematic bias of the model.

It should be noted that the interpretation of the Pearson correlation coefficient in event-based analyses involving extreme PM₁₀ concentrations requires caution, as the metric can be sensitive to high-magnitude values and isolated concentration peaks. However, in the present study, such extreme values are not considered statistical anomalies or spurious outliers but rather constitute the primary focus of the analysis, since the selected simulation periods correspond to intense dust episodes and high-impact PM₁₀ events. Consequently, the occurrence of large concentration values and strong temporal variability is an inherent characteristic of the dataset. For this reason, the correlation coefficient is interpreted together with complementary statistical metrics, including RMSE and MBE, as well as through the analysis of temporal series and spatial PM₁₀ distributions, providing a more robust assessment of model performance during extreme dust conditions.

Finally, 24 h moving averages were also calculated to evaluate the ability of the model to capture daily dust variability, which is critical for air quality management and forecasting. The capability of the newly developed datasets to reproduce dust events in WRF-Chem simulations was assessed by computing metrics for several dust intrusion episodes spanning multiple seasons.

Figures  and  show the new high-resolution soil erodibility fields for domains d01 and d03, respectively. In addition, Table  summarizes the spatially averaged erodibility values for both domains, together with the total number of valid (non-zero) erodible grid cells for each dataset.

The EROD-HR dataset shows notable differences relative to the CTRL configuration. For domain d01, the average erodibility of the clay-silt classes, which dominate dust emissions in GOCART, is 0.0111 when all grid points are considered and 0.0474 when only source points are included. These values represent differences of -0.53% and -1.45%, respectively, relative to the original dataset. Despite the small change in domain-averaged erodibility, the spatial distribution differs markedly from the default field, as commented in Sect. , with higher erodibility values concentrated in specific areas (Fig. ). This redistribution, together with a reduction in the number of source grid cells from 20 620 to 18 251 (-11.5%), indicates that while the overall erosion potential of the domain remains nearly unchanged, the location of dust-emitting areas is substantially modified.

A more pronounced effect is observed over the RM (domain d03), where the number of source grid cells decreases from 18 276 (58.1 % of the total) to 4603 (14.6 %). In contrast, the mean erodibility of these source points increases to 0.023, approximately one order of magnitude larger than in the original dataset. As shown in Fig. , the default CTRL configuration exhibits lower erodibility values distributed over broad areas, whereas the EROD-HR dataset produces a more heterogeneous pattern with localized high-erodibility regions. Nevertheless, the domain-averaged erodibility considering all grid points increases by only 0.20 %, suggesting that the total dust uptake potential in domain d03 remains largely unchanged.

The behavior of the SOILHD dataset differs substantially from that of EROD-HR. As previously discussed in Sect. , the SOILHD formulation introduces a physically based redistribution of erodible fractions by explicitly incorporating soil texture information. This results in a modification of the spatial organization of source areas while preserving the overall magnitude of domain-integrated erodibility. In particular, the inclusion of soil-type heterogeneity leads to a more realistic representation of fine-particle availability, while maintaining consistency with the baseline magnitude constraints imposed during the rescaling procedure (see Sect. ).

For domain d01, the number of source grid cells increases to 22 045 and 22 020 for the silt and clay classes, respectively, representing increments of 6.9 % and 6.8 %. In contrast, the mean erodibility of these source points decreases to 0.0199 (silt) and 0.0185 (clay). Despite this reduction, the relative contributions of the silt and clay fractions to total erodibility amount to 25.3 % and 23.5 %, respectively, indicating that the assumption of an average sand content of 50 % (i.e., a 50–25–25 % sand-silt-clay composition) remains reasonable for domain d01.

In contrast, the 50–25–25 % soil texture assumption is not suitable for the RM (domain d03). The SOILHD configuration shows a broader spatial distribution of erodibility than EROD-HR, with values comparable to the CTRL configuration but confined to more localized areas at higher spatial resolution. The total number of source points remains close to that of the original dataset (17 455, corresponding to a reduction of 4.7 %). This behavior reflects a clear shift in soil composition within domain d03, where fine particles play a more relevant role than initially assumed, leading to a more heterogeneous and spatially structured erodibility field.

This is consistent with the statistical breakdown reported in Table , which shows that the silt fraction accounts for approximately 40 % of total erodibility (0.0066), while the clay fraction contributes about 28 % (0.0046). Accordingly, a soil texture distribution of 30–40–30 (sand-silt-clay) appears more representative for this region. Consequently, the commonly assumed 50–25–25 % distribution likely underestimates the availability of fine erodible particles in the RM, and thus the potential for dust emission.

The difference in the number of source points between the EROD-HR and SOILHD datasets highlights limitations associated with computing the source strength function S at local-scale resolution. In highly erodible regions of the Sahara Desert, this approach can yield very low or near-zero values of S, as illustrated in Fig. . The inclusion of soil texture information in SOILHD partially compensates for this effect by increasing the number of active source cells while preserving realistic erodibility magnitudes.

Changes in the relative contribution of dust emission by size bin are illustrated in Fig. . Emissions in the CTRL and EROD-HR experiments are nearly identical. In contrast, SOILHD exhibits a marked reduction in emissions from bin 1 (0–1 µm, clay), decreasing from 8.9 % in CTRL to 6.3 % (a reduction of 28.6 %), accompanied by a slight increase in emissions from bins 2 to 4 (silt). The coarsest size class (bin 5) is excluded from the analysis, as it comprises particles with effective diameters larger than 12 µm, exceeding the upper size threshold of PM₁₀.

Finally, average erodibility values were also computed for the previously published erodibility dataset described by , which integrates remote sensing products and soil properties to represent the spatial and temporal variability of dust source strength. For domain d01, the resulting mean erodibility values are 0.035 when averaged over all grid cells and 0.166 when considering only source points, indicating a spatially homogeneous erodibility pattern over the Sahara Desert. These values are approximately three times higher than those obtained with the CTRL dataset.

For domain d03, the mean erodibility at source points is 0.02, comparable to the EROD-HR dataset, but the number of active dust source cells is limited to only 784, corresponding to 2.5 % of the total grid points. As a result, the domain-averaged erodibility is very low (0.0005). This configuration favors dust contributions dominated by transport from the parent domain, with negligible local dust uptake, which may lead to overestimated PM₁₀ concentrations associated with convective transport (see Sect. ). Such a behavior appears inconsistent with the known characteristics of the RM, which exhibits semi-arid conditions, low precipitation, and widespread areas of sparsely vegetated soils.

Additionally, the strong dependence of this kind of dataset on remote sensing retrievals may result in uncertainties and misvalues. For example, higher erodibility values are observed over the Alps and the Pyrenees, two mountainous regions in central Europe and northern IP, respectively, characterized by rocky, snowy areas with negligible sand content, and therefore null dust uptake potential.

Spatially averaged PM₁₀ profiles were computed for domains d01 and d03 during the August 2025 event to examine the vertical distribution of PM₁₀ (Fig. ). The motivation for analyzing the vertical structure is to assess whether the differences in surface erodibility translate not only into changes in near-surface concentrations but also into modifications of the vertical redistribution of dust within the atmospheric column. In particular, changes in the erodibility field directly affect dust emission fluxes at the surface, which in turn influence the vertical mixing and uplift of particles through turbulent transport and boundary layer processes. For domain d01, the vertical PM₁₀ profile shows a significant reduction in concentrations in the EROD-HR experiment relative to the CTRL one. This reduction is more pronounced in the SOILHD configuration, particularly near the surface, where the mean PM₁₀ difference with respect to CTRL reaches -63.4 µg m⁻³.

Over the RM (domain d03), the vertical PM₁₀ profiles obtained with CTRL and EROD-HR are nearly identical. In contrast, the SOILHD experiment exhibits lower PM₁₀ concentrations than CTRL at all vertical levels, with a maximum difference of 47.0 µg m⁻³ at model level 11 (approximately 750 hPa). This level also corresponds to the maximum PM₁₀ concentrations, indicating that advective transport plays a dominant role in the vertical dust distribution over the region. As a result, modifications in the erodibility distribution of the parent domain propagate into significant differences in the simulated dust burden within the inner domain.

At the surface level, the mean PM₁₀ difference between SOILHD and CTRL amounts to 7.9 µg m⁻³. Comparable behavior is observed for the remaining simulated periods. These results indicate that changes in soil mass fraction representation introduced by SOILHD affect both dust uplift and transport processes, leading to a modified vertical distribution of PM₁₀.

Additional simulations were performed using the erodibility dataset developed by as the soil erodibility input (LI2023 experiment). The resulting vertical PM₁₀ distributions for domains d01 and d03 indicate substantially higher concentrations than in all other experiments, with values that are two to three times larger than those obtained with the CTRL, EROD-HR, and SOILHD configurations. In domain d03, maximum PM₁₀ concentrations exceed 250 µg m⁻³, with surface values around 100 µg m⁻³.

Similar behavior is observed for other simulated periods. For instance, during the 2022-07 event, a maximum PM₁₀ concentration of 149.2 µg m⁻³ is obtained with LI2023, whereas the CTRL, EROD-HR, and SOILHD experiments yield maxima ranging between 40.9 and 77.8 µg m⁻³. This contrast suggests that the LI2023 dataset produces excessive dust emissions in the parent domain (d01) while generating very limited local dust uptake over the RM. Such a configuration is difficult to reconcile with the semi-arid conditions and the widespread presence of sparsely vegetated soils characteristic of the region.

To provide a spatially explicit assessment of how the different erodibility configurations affect the simulated PM₁₀ distribution, a domain-scale relative difference analysis was included. Figure  shows the temporal mean PM₁₀ fields for the CTRL experiment (top left) during the 2025-08 period, together with the relative differences induced by the EROD-HR, SOILHD, and LI2023 configurations over domain d03, with the SINQLAIR stations indicated in red.

Overall, the LI2023 dataset produces a clear and systematic overestimation of PM₁₀ across the entire domain. Relative differences exceed +100% in most of the region, indicating a strong enhancement of dust loading with respect to the CTRL experiment. This behavior is consistent with the erodibility statistics in Table , which show exceptionally high erodibility values over domain d01 and a highly limited number of active source cells in domain d03. It is also consistent with the results shown in Fig. , where LI2023 exhibits substantially higher PM₁₀ concentrations at all vertical levels. This combination leads to enhanced advective transport from the parent domain and an overall amplification of dust concentrations in the inner domain. Based on this and all the previous results, the LI2023 experiment was excluded from the comparison with ground-based observations presented in the following section.

In contrast, the EROD-HR configuration produces PM₁₀ fields that are very similar to those of the CTRL experiment. Relative differences are generally small and mostly below 10 %–20 % over most of the domain, with localized maxima approaching 40 % in a few specific areas. These differences are spatially constrained and do not modify the overall structure of the PM₁₀ field, indicating that the increase in horizontal resolution of the erodibility dataset alone does not significantly modify the modeled PM₁₀ concentrations at the surface.

Finally, the SOILHD experiment shows a more complex spatial response. On average, it yields lower PM₁₀ concentrations than CTRL, with widespread negative differences of approximately -10% to -25% across large parts of the domain. However, localized regions also exhibit positive deviations of up to approximately 10 %, indicating that the impact of SOILHD is not uniform in space. This spatial heterogeneity reflects the introduction of soil texture information and the redistribution of erodible fractions, which modifies both the intensity and the location of dust source regions. As a result, SOILHD not only reduces overall dust loading in many areas but also enhances emissions in specific localized hotspots where soil composition favors higher erodibility.

Surface PM₁₀ concentrations from all simulations were evaluated against observations from the SINQLAIR regional air quality monitoring network within the RM domain (d03). The evaluation was conducted across five independent simulation periods representing distinct synoptic and emission regimes, ranging from large-scale Saharan dust outbreaks to local uplift and convectively driven events. This multi-event framework allows assessment not only of the overall performance of the different erodibility configurations but also of their robustness under contrasting interannual meteorological conditions and spatial emission patterns. Figure  summarizes the distribution of correlation coefficients (left) and relative RMSE values (right) obtained for the modeled PM₁₀ time series at all monitoring stations and simulation periods.

Substantial variability is observed across the analyzed periods, reflecting the strong dependence of model performance on the dominant dust generation and transport mechanisms. Nevertheless, a consistent behavior emerges across most simulations: the inclusion of the high-resolution erodibility datasets, and particularly the SOILHD configuration, generally improves agreement with observations relative to CTRL. Overall, SOILHD yields the most stable behavior across periods, with systematically higher median correlations and lower RMSE dispersion. This indicates that the improvements introduced by the soil-texture-dependent erodibility formulation are not restricted to a single event type but remain robust under different synoptic conditions and emission regimes.

The magnitude of the improvement, however, varies substantially depending on the nature of the event. During intense dust outbreaks dominated by long-range Saharan transport, such as the 2022-03 episode, all simulations exhibit very large RMSE values due to the extreme PM₁₀ concentrations reached during the event. In contrast, for periods more strongly influenced by local uplift and mesoscale variability (e.g., 2022-07 and 2025-08), absolute RMSE values are lower, and differences between experiments become less pronounced, although still significant. In these local-scale situations, SOILHD produces relative RMSE reductions of approximately 12.0 % and 21.0 %, respectively, compared to CTRL.

A marked spatial variability is also evident among stations. Inland stations generally show a clearer separation between experiments, with SOILHD consistently reducing both RMSE and bias relative to CTRL and EROD-HR. Conversely, coastal stations tend to exhibit greater variability and weaker sensitivity to the erodibility configuration. This behavior is likely associated with the combined influence of marine aerosol contributions, local recirculation processes, and anthropogenic sources not represented in the present simulations.

The 2022-03 event represents the most extreme episode analyzed in this study, characterized by persistent and exceptionally high PM₁₀ concentrations across the entire IP and specifically over the RM. Under these conditions, the improvement achieved with SOILHD becomes particularly evident. The average correlation coefficient increases from 0.43 in CTRL to 0.57 in SOILHD, with a maximum station correlation of 0.69 compared to 0.57 in CTRL. At the same time, the average RMSE decreases from 358.2 % to 191.3 %, corresponding to a relative RMSE reduction of 166.9 %. Mean bias error (MBE) also decreases substantially, from 181.4 to 74.1 µg m⁻³ (105.6 µg m⁻³ for EROD-HR), further demonstrating the improved representation of dust emissions provided by SOILHD during intense transport-driven episodes.

In contrast, the 2022-07 period exhibits the largest inter-station variability in correlation values. CTRL correlations range from values close to zero (0.07) up to 0.61, whereas EROD-HR and SOILHD slightly reduce this dispersion, yielding ranges of 0.27–0.68 and 0.21–0.68, respectively. This behavior suggests that higher-resolution erodibility fields improve the representation of localized emission structures and temporal PM₁₀ evolution under weak synoptic forcing conditions. During this event, CTRL underestimates PM₁₀ concentrations at six of the eight stations, with an average MBE of -4.00 µg m⁻³ and RMSE values frequently exceeding 20 µg m⁻³. In addition, only 50 % of the stations exhibit statistically significant correlations (p-value <0.01). The EROD-HR experiment slightly reduces RMSE values but tends to maintain the overall underestimation pattern, whereas SOILHD further improves the temporal structure and magnitude of the simulated dust peaks.

The 2025-04 period represents an intermediate situation combining convective dust transport from northern Algeria and Morocco with local uplift processes along the eastern Iberian coast. Correlation values remain very similar among experiments, with SOILHD showing a marginally higher average correlation (0.75) relative to CTRL (0.71) and EROD-HR (0.72). However, substantial differences emerge in RMSE values. In particular, EROD-HR produces unrealistically high PM₁₀ concentrations, yielding an average RMSE of 121.6 µg m⁻³, the highest among all experiments and periods analyzed in relation to the average PM₁₀ observations (rRMSE of 485.4 %). This behavior is likely associated with highly localized regions of excessive erodibility that become strongly activated under intense wind conditions. In contrast, SOILHD provides the lowest average RMSE (44.0 µg m⁻³), indicating that the inclusion of soil-texture heterogeneity moderates unrealistic localized emissions while preserving the spatial variability of dust uplift.

The 2025-08 event, which is examined in greater detail later in this section, yields the best overall performance for SOILHD. This period exhibits the highest average and maximum correlation values among all simulations, together with substantially reduced RMSE values relative to CTRL and EROD-HR. These results further highlight the capability of SOILHD to improve the representation of local-scale dust uplift and convectively driven transport situations.

Finally, the 2025-11 event is characterized by a strong Atlantic low-pressure system promoting intense southwesterly flow and northward dust transport from North Africa toward the IP. During this period, SOILHD exhibits a more consistent correlation across stations. Although the mean correlation is slightly higher for EROD-HR (0.70 compared to 0.68 for SOILHD), the median value is higher for SOILHD (0.70 versus 0.69), indicating that the apparent advantage of EROD-HR is largely driven by a single outlier station (LOR, 0.84). This suggests a more homogeneous performance of SOILHD across the monitoring network, whereas EROD-HR shows greater sensitivity to site-specific behavior. RMSE values slightly increase on average with SOILHD, from 62.4 to 74.5 µg m⁻³. This constitutes the only analyzed period in which both RMSE and MBE increase in SOILHD relative to CTRL, a behavior also partially reproduced by EROD-HR. This result likely reflects the sensitivity of long-range transport events to the precise spatial redistribution of erodible areas introduced by the high-resolution datasets. Under these conditions, the modified source regions incorporated in SOILHD may enhance emissions in areas with stronger effective erodibility than those represented in the default configuration.

Overall, SOILHD yields the highest average correlation values in three out of the five analyzed periods, while producing nearly identical values to EROD-HR during the remaining two events (2022-07 and 2025-04). In addition, SOILHD provides the lowest RMSE values in four out of five periods, the only exception being 2025-11. Both EROD-HR and SOILHD consistently produce statistically significant correlations (p-value <0.01), confirming the robustness of the improvements introduced by the high-resolution erodibility formulations. To further complement the statistical evaluation across periods and experiments, a Taylor diagram summarizing the spatially averaged performance metrics (correlation, standard deviation, and RMSE) for each simulation period has been included in the Appendix (Fig. ). This provides a compact visual synthesis of model skill and its variability across the different interannual regimes analyzed in this study.

The August 2025 case was examined in greater detail, as it represents a summer dust episode characterized by the combined influence of local dust uplift and convective transport, a situation in which the physical configuration of the erodibility field is particularly influential. Figure  presents surface PM₁₀ observational (OBS) and modeled (CTRL, EROD-HR, and SOILHD) time series (dashed lines) for the 2025-08 period, obtained by bilinear interpolation of model output to the geographic coordinates of the SINQLAIR monitoring stations. In addition, 24 h moving averages (24MA) are shown for all series (solid lines), both to improve clarity and visualization and because 24 h averaging is a standard metric in air quality analysis and forecasting. Complementary PM₁₀ boxplots summarizing the distribution at each station and for every dataset are also included. For a more detailed inspection of individual station time series, Fig.  is provided.

Overall, the SOILHD configuration provides the best agreement with observations at most monitoring stations, particularly during the main dust event of 27–28 August, when it successfully captures peak PM₁₀ concentrations at multiple sites. From 26 August onwards, the 24 h moving-average (24MA) observational series remains within the envelope defined by the individual station time series simulated with SOILHD, indicating that this configuration produces PM₁₀ concentrations which are representative of the observed inter-station variability. In addition, SOILHD closely matches the observations, reproducing both mean and peak PM₁₀ values at most locations.

In contrast, both the CTRL and EROD-HR experiments substantially overestimate PM₁₀ concentrations across all stations, particularly during high-concentration episodes. For instance, at the San Basilio (SBA) station, CTRL and EROD-HR simulations reach peak values exceeding 250 µg m⁻³, whereas the observed maximum is approximately 150 µg m⁻³.

It is worth noting that all simulations in this study only include mineral dust emissions, while anthropogenic PM₁₀ sources (e.g., NO_x, SO₂, BC, and secondary aerosol formation pathways) are not explicitly represented. Consequently, part of the persistent background PM₁₀ concentrations observed at the monitoring stations is not reproduced by the model. This behavior is evident in both Figs.  and , where observed PM₁₀ values remain well above near-zero levels outside the main dust peaks, indicating a non-negligible anthropogenic contribution, particularly at inland and industrially influenced locations such as LOR. The absence of these background sources may partly explain the systematic underestimation found in some periods and stations, and should therefore be considered when interpreting the absolute PM₁₀ biases. Under these conditions, a moderate negative bias in the simulations is physically plausible and, to some extent, expected.

Although all configurations yield similar correlation coefficients (0.89), the SOILHD experiment exhibits a substantially lower RMSE (15.7 µg m⁻³), highlighting its superior quantitative performance. A closer examination of the individual time series at each monitoring station (Table ) confirms that the SOILHD configuration provides the best overall performance. SOILHD yields lower RMSE values at 7 out of 8 stations, with an average reduction of 10.9 µg m⁻³ relative to the CTRL experiment. Although all configurations exhibit high correlation coefficients at most sites, SOILHD attains the highest correlation at five of the eight stations, with an average value of 0.77.

When switching to the new high-resolution erodibility datasets, near-coastal stations (ALU, ALJ, ESC, and MOM) exhibit a weaker, yet still statistically significant, response in terms of correlation, together with a marked improvement in RMSE relative to observations, with the exception of ESC. At these sites, all model configurations tend to overestimate PM₁₀ concentrations, suggesting an overestimation of local dust source strength and/or excessive recirculation within the model domain.

The ESC station represents a particular case, as its observed PM₁₀ variability is substantially higher than at any other site. This behavior may be partly explained by the contribution of sea-salt aerosols to PM₁₀ levels and by the proximity of the station to the Escombreras combined-cycle power plant, a known source of black carbon, nitrates, sulfates, and other anthropogenic pollutants not explicitly represented in the simulations. Nevertheless, the model is able to partially reproduce the observed variability at ESC, particularly in the EROD-HR and SOILHD experiments. These configurations also capture the magnitude of several PM₁₀ peaks, most notably during the dust outbreak of 27–28 August. Among the simulations, the modeled time series transitions from a slight overestimation in the SOILHD experiment to a more pronounced overestimation in CTRL and EROD-HR.

For inland stations (ALC, CAR, LOR, and SBA), CTRL and EROD-HR exhibit nearly identical behavior, substantially overestimating observed PM₁₀ concentrations, especially during the peak associated with the 27–28 August dust event. In contrast, the SOILHD experiment shows a remarkable agreement with observations, accurately reproducing both the timing and magnitude of the PM₁₀ maximum. For instance, at Alcantarilla (ALC), SOILHD simulates a maximum concentration of 165.1 µg m⁻³ on 27 August, in close agreement with the observed value of 167.6 µg m⁻³ (relative error of 1.5 %). In comparison, CTRL and EROD-HR yield maxima of 266.6 and 247.1 µg m⁻³, corresponding to errors of 59.1 % and 47.4 %, respectively. The only notable exception to the overall pattern is observed at the Lorca (LOR) station, where both CTRL and EROD-HR simulations more closely capture the observed peak PM₁₀ concentrations. In contrast, SOILHD consistently exhibits a negative bias throughout the study period. This discrepancy is likely attributable to the strong influence of anthropogenic pollution at LOR, which is not represented in the current simulations. Local industrial activity and traffic emissions, each with distinct diurnal cycles, may contribute an additional PM₁₀ background that is absent from the model, resulting in systematic underestimation in highly polluted areas.

To further illustrate these dynamics, Fig.  presents the time series of PM₁₀ and accumulated dust for all eight stations during the 2025-08 period. For coastal stations (ALU, ALJ, MOM, and ESC), it is evident that the accumulated dust concentrations alone cannot account for the modeled PM₁₀ levels, indicating a substantial contribution from marine salt to the natural-origin PM₁₀ fraction (the only component represented in the present simulations). Conversely, at some stations, namely ALC, ALJ, MOM, LOR, ESC, and SBA, SOILHD systematically underestimates the observed PM₁₀ concentrations. The observed deviations can be attributed to a combination of external PM₁₀ sources specific to each location, such as heavy traffic (ALC, MOM, LOR, SBA), industrial emissions (ALC, MOM, ESC), and agricultural activities such as soil tilling (ALJ).

Based on the discrepancy between observations and the SOILHD experiment, which performs well in low-pollution environments, this anthropogenic contribution can be estimated to range between approximately 25 and 50 µg m⁻³. This interpretation is further supported by results at Caravaca (CAR), a station located in the northwestern part of the RM, where anthropogenic influences are minimal. At this site, SOILHD clearly provides the best agreement with observations. It is important to note that the primary objective of this study is to evaluate the performance of the new erodibility database, rather than to develop an operational air-quality forecasting system. The incorporation of anthropogenic emission inventories is therefore beyond the scope of the present work, but will be addressed in future simulations.