Review of “Improving Fine Mineral Dust Representation from  the Surface to the Column in GEOS-Chem 14.4.1”

The manuscript titled GMD_2025_Improving Fine Mineral Dust Representation from the Surface to the Column in GEOS-Chem 14.4.1 describes the implementation of the latest dust emission scheme in the GEOS-Chem High Performance configuration (version 14.4). The authors emphasize the need to reconcile the modeled geometric diameter with in-situ measurements that are based on aerodynamic diameter. They also provide a comprehensive, step-by-step approach to refining and evaluating the model, effectively isolating uncertainties by region. The manuscript is well-organized, and the thoroughness of the performance evaluations is particularly commendable. Overall, I am impressed by the clarity and depth of this work, and I recommend acceptance with minor revisions.

Specific comments:

1. Dust Optical Properties in GEOS-Chem: It would be helpful to include a concise summary of the dust optical properties used in your GEOS-Chem configuration. You mention that the model employs the improved dust optical properties from Singh et al. (2024). Could you clarify whether the aspect ratios of the various dust bins in your model are consistent with those in Singh et al. (2024)
2. Size Distribution Comparisons: Your analysis would be more compelling if you compared the model’s size distribution against in-situ measurements, including AERONET data and other publicly available datasets (e.g., doi.org/10.5194/essd-16-4995-2024). Such comparisons would provide additional evidence that your updates to the dust emission scheme accurately capture real-world size distributions.

Technical corrections:

In your figures, the circles appear to lack any visible color fill. Consider adjusting the visualization to ensure the colors are clear and distinguishable.

Reporting “AW 2.3 ± 5.7” can be misleading because concentrations and emissions generally cannot be negative. Replacing the AW values with confidence intervals or alternative statistical measures might be more informative and intuitive for readers.

"Improving Fine Mineral Dust Representation from the Surface to the Column in GEOS-Chem 14.4.1" by Zhang et al. is a model development paper that evaluates the dust aerosol cycle in various configurations of the GEOS-Chem Chemical Transport Model. The paper is well-organized, although some key information is occasionally missing, which makes the paper difficult to fully evaluate. Most of my comments are requests for clarification. Overall, I found the paper illuminating and worth reading, especially as a dust scientist working with a different ESM. I am recommending that the paper be accepted subject to major revision. The authors can contact me at ron.l.miller@nasa.gov if they have any questions about my review.

1. Some aspects of the model are incompletely documented. For example, the authors note that emission is calculated using winds from the high-resolution GEOS-FP Forward Processing model used for weather forecasting, whose horizontal resolution (0.25° latitude by 0.3125° longitude) is high compared to typical global dust models as well as the resolution of the GEOS-Chem transport model used here (closer to 2 degrees). However, I could not find how often emission is calculated: eight-times daily? Also, given that many studies nudge toward the GEOS MERRA2 product, could the authors tell us a little about the relation of the winds of the Forward Processing model and MERRA-2?

Similarly, the optical properties of dust particles are not fully described. The study by Singh et al. (2024) is cited ('updated optical properties for aspherical mineral dust'). I think this means that they are modeling dust as spheroids for scattering calculations, but they should clarify this as well as cite their source for the dust index of refraction. Singh et al., 2024 cite Tegen and Lacis, who use an index from measurements by Patterson et al. (1977). However, later Singh et al. cite the index compiled by Sinyuk et al. (2003). Please clarify how the index of refraction and optical properties were prescribed.

Patterson, E. M., D. A. Gillette, and B. H. Stockton (1977), Complex index of refraction between 300 and 700 nm for Saharan aerosols, J. Geophys.  Res., 82, 3153–3160.

Sinyuk, A., O. Torres, and O. Dubovik, Combined use of satellite and surface observations to infer the imaginary part of refractive index of Saharan dust, Geophys. Res. Lett., 30(2), 1081, doi:10.1029/ 2002GL016189, 2003.

Removal processes could also be described in greater detail. Figure 9 shows the dry deposition velocity as a continuous function of particle size, even though it is prescribed in the model afor discrete bins.  The same is true for the new washout parameterization. Figure 9 would be more useful if the authors plotted the discrete values of both the deposition speed and washout rates for each bin for both the 4 and 7-bin versions of the model. (i.e. replace the continuously varying washout rate with 7 discrete values for comparison to the two values used in the default model that are currently identified by an orange dashed line.)

2. The authors adjust the global emission so that spatial variations in model AOD match those from the annual average Deep Blue AOD retrievals, according to a regression criterion (line 336). First, I cannot tell whether the base model is also calibrated in the same way.

Second, a limitation of this method is that the model AOD depends not only upon dust but all the other aerosol types computed by GEOS-Chem.  Thus, tuning the model to match Deep Blue may compensate for errors in these other fields and introduce biases into the dust AOD and the surface concentration.  The model improvements described by the authors may not be entirely addressing limitations in the dust model, but rather biases in the other constituents. It is probably beyond the scope of this paper to address non-dust biases in AOD, but this uncertainty should be given greater emphasis in the method description and conclusions.

3. Another uncertainty that is not addressed is the temporal mismatch of the simulation period and measurements for evaluation. All the simulations are for the year 2018. In contrast, the CALIOP retrievals are for 2007-2021 while the SPARTAN dust PM2.5 network spans the five years between 2019 and 2023. (Some SPARTAN stations are based on as few as 10 measurements, which means that their annual average is subject to a potentially large sampling uncertainty.) This mismatch is partly the result of data availability, but it should at least be acknowledged in the conclusions as part of a fuller description of uncertainty.

4. I like the careful comparison by the authors of the SPARTAN surface concentration measurements that are characterized by aerodynamic diameter to the geometric diameter used in the model. They note that even larger surface biases would result without this correction.

This raises a question about the washout parameterization: (line 507, eq. 9). Is the dependence of washout rate to dust particle size derived or fitted assuming spherical particles? The authors are careful to use more irregular shapes when calculating optical properties (spheroids) and comparing to concentration measurements (ellipsoids?). Have the authors accounted for non-spherical shapes in their washout calculation? Ellipsoids and spheroids might have a greater chance of washout, comparable to spheres with larger diameter.

5. There are a few ad hoc assumptions that should be given more emphasis.  The authors '[apply] a regional scaling factor of 0.6 over the Sahara to reduce its emissions' (line 367). This reduction seems arbitrary, and it has a big impact upon model behavior given the global importance of the North African source. Is there a physical basis for this? This rescaling should be acknowledged in the conclusions.

Similarly, the dependence of the wind speed threshold for emission upon soil moisture is treated somewhat arbitrarily. In particular, the authors reduce by half the soil moisture from the top layer of the GEOS-Chem land model before applying it to the reduction of the wind speed threshold for emission. In support, the authors cite Darmenova et al. (2009), along with Wu et al. (2022); the latter justify their own reduction by citing Darmeonova's doctoral thesis. Darmenova et al.  point out that measurements linking the wind speed threshold and soil moisture are based upon the upper 1-2 cm. They note the coarser resolution of climate model soils, while citing measurements at different locations to argue that moisture near the surface should be less than moisture integrated over a deeper layer. The issue is that following precipitation, drying will start at the surface and propagate downward so that the upper 1-2 cm dry out first, before the deeper top layer in GEOS-Chem. Thus, as noted by Darmenova, the GEOS-Chem uppermost layer has too much inertia compared to the surface moisture that is observed to limit dust emission. The authors should present observations that show the relation of soil moisture at 1-2 cm compared to the 5-cm layer integral. Otherwise, they should acknowledge that their rescaling of the 5-cm GEOS-Chem value is a source of uncertainty.

6. The Conclusions are brief and could be expanded to acknowledge uncertainties and the implications of results. For example, could you choose the dry deposition speed and the washout rate of DST1 in the 4-bin model so that it matches the removal rates of the 7-bin model? This would give you the improvement seen for 7 bins but with greater computational efficiency.

Also, the model is evaluated using annual average observations, but seasonal biases are relegated to the appendix. It would be helpful to have some discussion of seasonal model behavior including biases in the main article.

Minor Comments:

73 'predicted spatial distribution *of emission*'

90 'an overestimation of fine dust (Kok, 2011; Kok et al., 2017)' see also Cakmur et al 2006

Cakmur, R.V., R.L. Miller, J.P. Perlwitz, I.V. Geogdzhayev, P. Ginoux, D. Koch, K.E. Kohfeld, I. Tegen, and C.S. Zender, 2006: Constraining the global dust emission and load by minimizing the difference between the model and observations. J. Geophys. Res., 111, D06207, doi:10.1029/2005JD005791.

96 'especially over size ranges with rapid variation in processes' Does 'rapid' refer to variations with respect to size?

140 '10 samples for the 5-year' Have you tested a higher threshold for minimum measurement number? You are trying to resolve an annual average with only 10 samples over five years, which could be a problem if the measurements all occur within a particular season.

158 'are computed with relative humidity dependent aerosol size distributions' Does the model calculate deliquescence of dust particles and its effect upon optical properties?

166 'standard wet deposition scheme includes scavenging in convective updrafts, and in-cloud and below-cloud scavenging from precipitation.'  What controls the rate of scavenging in convective updrafts? Is it different from the precipitation-rate dependence in stratiform clouds?

198 'where 𝑓 is the clay content in the top soil layer and a global constant value of 0.2 is used to reduce excessive sensitivity of dust emission fluxes to 𝑓' This is unclear. Is fclay set equal to 0.2 or multiplied by 0.2?

203 'Brittle Fragmentation Theory (Kok, 2011) with parameter values optimized using dust observations from the Interagency Monitoring of Protected Visual Environments (IMPROVE) ground-based monitoring network in the United States' Is this correct? Zhang et al. (2013) say that the sidecrack propagation length of 8 um is taken from Zhao et al. (2010), who fitted to measurements of dust over North Africa during the DABEX field campaign.

237 'The mass fraction of each simulated dust size bin to the total fine dust mass concentrations can be calculated by the integration of the dust size distribution of Equation (4) with the 𝜆 value of 12 μm of the default PSD used in the GEOS-Chem (GC PSD)' I'm confused, On the previous page (line 210), you quote 8 um as the default value.  Second, eq. 4 describes the PSD of emission, but you are using it here to represent the PSD of load.  Is this described correctly?

280 'The simulated vertical profile shows excellent agreement against the 15-year (2007 to 2021)' To me, Figure 3 shows a consistent low bias of model dust in the lowest kilometer, with a corresponding high bias above? (This underestimate seems odd since model surface concentration is overestimated in comparison to SPARTAN: Figure 2).

336 'The total global annual source strength for each sensitivity simulation is scaled to achieve unity slope versus Deep Blue AOD (Figure A1) over major dust source regions.' Is this also true for the BASE experiment? Is dust AOD (DAOD) calculated separately? What is the global DAOD for each experiment?

342 'Regression equations are calculated using reduced-major-axis linear regression' Could you give a brief description of reduced-major-axis linear regression and how it compares to the standard technique? How are you calculating the uncertainties for both the model and observations?

398 'In addition, we reduce the sensitivity of dust emissions to clay content by eliminating the multiplication of the capped clay content 𝑓′ Wouldn't the removal of a factor less than one increase the sensitivity to clay? On line 198, you write that the addition of this factor is intended to reduce sensitivity.

400 'Soil wetness is taken from the parent meteorological inputs of GEOS-FP.' How reliable is the GEOS-FP soil wetness?

422 'Eliminating the multiplication of the capped clay content of 𝑓′ reduces the dust emission sensitivity to the clay 𝑐𝑙𝑎𝑦 content, increasing emissions' see comment on line 398.

435 'with improvements to the relative regional magnitude of dust across the Sahara, Middle East and Asia.' Is this true? According to the R2, NMD, NRMSD metrics in Figure 6, all the sensitivity experiments perform worse over Asia compared to the BASE simulation.

Figure 6: why for each region does the number of observations vary according to the experiment? Is this because the number of locations where AODdust/AOD exceeds 0.5 varies across the sensitivity experiment? If so, perhaps note this in the caption?

480 'we adopt the Kok PSD' This is confusing. You already adopt the Kok PSD in your BASE experiment. What you are changing here is the sidecrack propagation length estimated by Kok. I suggest identifying this experiment by the lambda value rather than 'Kok'.

504 'varying by 3 orders of magnitude for 505 diameter ranging from 1 to 10 μm' Wouldn't the contrast in the DST1 bin between 0.1 and 1 um be more relevant to fine dust?

540 Please identify the 'ground-observations'. 

549 'The simulated total column AOD would be underestimated by 14% compared to AERONET...' Is emission of the low-res model rescaled so that regression of model AOD versus Deep Blue has unity slope? More generally, the high performance version has been subjected to numerous adjustments to improve agreement with measurements and retrievals, so the low-res version is at a disadvantage.

610 'overall consistency in the vertical shape' see comment on line 280

619 'the Kok particle size distribution (PSD; Kok, 2011) better represents the mass fraction of fine dust measured during the Fennec field campaign over Northern Africa than the default PSD'. I agree that restoring the Kok value of lambda to 12 um gives better results in GEOS-Chem, but I think the FENNEC results are being oversold as a justification for the change. The difference of the Kok and GC PSD in Figure 8 are small compared to the large spread of the FENNEC emitted size distribution.

The paper titled “Improving Fine Mineral Dust Representation from the Surface to the Column in GEOS-Chem 14.4.1” aims to enhance the simulation of fine dust in the GEOS-Chem model using the GCHP framework by updating dust-related parameters and implementing a new dust scheme. There are some interesting results, However, as an evaluation of global dust model performance, some important details are not clearly described, and I have significant concerns regarding the methodology, particularly the limited number of observational sites used for validation and the reliance on single annual mean values. With only ~16–26 globally distributed sites (from AERONET and SPARTAN) providing one annual value each for 2018, the statistical metrics (e.g., R², NMD, NRMSD) lack robustness and may not reliably represent global model performance. For meaningful dust model evaluation—especially in the context of dust prediction—it is more appropriate to assess the model’s ability to capture daily variability over key dust source regions rather than depending solely on globally averaged annual mean comparisons at sparse locations. I encourage the authors to address this limitation explicitly and consider incorporating higher temporal resolution evaluations and more regionally focused diagnostics to strengthen the assessment. I therefore recommend a major revision and suggest that the manuscript be resubmitted after the following comments and questions have been thoroughly addressed.

1. L133: It is unclear whether the SPARTAN data used in this study represent daily means, monthly means, or another temporal interval. Additionally, the temporal resolution of the AERONET data should be clearly described to ensure consistency and clarity in the model–observation comparison.
2. L239: In the default PDF used in the GEOS-Chem (GC PDS), the lambda is not 12, however 8, which is implemented by Zhang et al., 2013. Please double check the correct values and calculation.
3. L257 and Fig. 1: In GEOS-Chem and other models, PM2.5 dust is typically estimated using an empirical formula (e.g., bin1 + 0.38 × bin2) based on assumed size distributions within each bin. This approach differs from observational definitions of PM2.5, which are determined by instrument-specific inlet penetration efficiencies (e.g., WINS or SCC). In Figure 1, the authors appear to apply such instrument-based cut-off functions to define PM2.5 dust fractions (e.g., stating 67.6% of DST1 based on SCC 1.829), which differs from the standard model approach. However, it is unclear whether this correction is actually implemented in the model calculations or only used for comparison with observations. I suggest the authors clarify whether and how this method was applied, and if so, describe it explicitly in the methods section.
4. Section 3: Again, the method used by the authors to calculate PM2.5 dust could significantly affect the subsequent comparison with observations. It remains unclear whether the apparent overestimation of PM2.5 dust is due to differences in how PM2.5 dust concentration is defined or calculated in the model.
5. L286: The (AODdust/AOD ) >0.5 is based on GEOS-Chem model prediction or based on AERONET observation? The AOD comparison is also 2018 annual?
6. Figure 2. In relation to the previous point, there is also uncertainty in the ratio of AOD_dust to total AOD between the model and observational data. Many previous studies have shown that significant discrepancies in species ratios between model outputs and observations can strongly affect the reliability of model evaluation. However, the authors do not discuss this issue, nor do they provide any supporting evidence or analysis related to it.
7. Line 288: According to line 140, the study uses SPARTAN data from sites with at least 10 samples over the 5-year period (2019–2023) since the network began using XR. This raises concerns about the data quality used for comparison and validation in Figure 2 and the subsequent analysis. In particular, it is unclear how many valid observations are available for the PM2.5 dust comparison shown in Figure 2. Beyond the scatter plots, I strongly recommend that the authors include time series comparisons at several representative sites near dust source regions. Scatter plots between annual model simulations and observations may be misleading, especially if each site has only a few observations throughout the entire year.
8. Figure2: Figure 2 provides useful regression statistics such as NMD and NRMSD for comparing simulated AOD and surface PM2.5 dust against ground-based observations, the number of sites used in the analysis (approximately 16 for AERONET and 26 for SPARTAN) is relatively limited. Since each site contributes only a single annual value, the statistical robustness and global representativeness of these metrics may be limited. I suggest the authors discuss this limitation explicitly and consider including additional evaluations, such as seasonal or monthly comparisons if available, or using regional breakdowns to enhance the interpretability of the model performance.
9. Figure 3: The use of extinction profiles normalized by AOD introduces a scaling effect that can obscure absolute errors in the extinction values. While such normalization is helpful for assessing the relative vertical structure of aerosols, it is less effective for diagnosing biases that contribute to discrepancies in column vs. surface aerosol performance. A direct comparison of extinction values would more clearly indicate whether vertical profile inaccuracies — in both shape and magnitude — are responsible for the model’s surface vs. column performance differences. It would be helpful if the authors could clarify this choice and ideally provide a comparison using the unnormalized extinction as well.
10. L307: The authors concluded that the model overestimates PM2.5 dust based on a comparison with the 2018 annual mean from SPARTAN data using the scattering method. However, this conclusion is not comprehensive, as it does not incorporate other ground-based observations to assess the uncertainty of SPARTAN. Additionally, the authors do not provide information on data availability or temporal variability. I recommend validating this conclusion using additional PM2.5 dust datasets, such as AIRNOW or IMPROVE, to enhance the robustness of the assessment.
11. L315 and Figure 4: Could you please clarify the size distribution used for the default dust simulation (base) in GEOS-Chem? It would be helpful to know the specific aerosol size bins or the distribution function applied in the model for dust particles.
12. L328: What is the value of Lamda, 12? Did you compare with the result of using 8 in the default GC dust scheme?
13. Figure 6: In Figure 6, the number of validation grid points (N) differs across the dust experiments due to the use of a dust-region mask defined by AOD_dust/AOD > 0.5. I believe this approach is not entirely reasonable, as the dust source regions themselves should remain consistent when using the same dust source function. The variations in AOD across experiments are a result of changes in dust scheme parameters, not the geographic extent of dust source regions. By applying the AOD_dust/AOD > 0.5 threshold independently for each experiment, you may be artificially excluding regions where the AOD has been reduced due to your parameter changes. This not only alters the spatial domain being compared but also potentially removes part of the signal you aim to evaluate—namely, the impact of those parameter changes on AOD. As a result, the comparison across experiments is not entirely fair and may underestimate the full effect of the modified parameters on dust AOD. To ensure a fair comparison, I suggest to be based on the BASE run for the dust source region instead of Figure A2.
14. Figure 7: I would like to raise the same concern as using this validation methor to get conclusion that the performance are improve as : the number of sites used in the analysis (approximately 16 for AERONET and 26 for SPARTAN) is relatively limited. Since each site contributes only a single annual value, the statistical robustness and global representativeness of these metrics may be limited. I suggest the authors discuss this limitation explicitly and consider including additional evaluations, such as seasonal or monthly comparisons if available, or using regional breakdowns to enhance the interpretability of the model performance.
15. L437-442: Given the relatively small number of ground-based sites used in the evaluation (with only one annual mean value per site), I would caution against drawing strong conclusions from the changes observed in Figure 7. While the reduction in mean bias in surface PM2.5 dust is notable, the accompanying decrease in R² for both AOD and PM2.5 suggests a potential loss in spatial agreement. This appears to contradict the claim that AOD skill "remains comparable" to the base simulation. I recommend that the authors either provide statistical support for this conclusion (e.g., uncertainty estimates or significance testing), or qualify the statement to reflect the limitations of the evaluation dataset.
16. Figure 11: Although the reduced-major-axis linear regression slope for PM2.5 dust against SPARTAN observations is further reduced from 1.53 (Figure 8) to 1.44, and the NMD remains comparable, I am concerned that this apparent improvement in surface PM2.5 dust may come at the cost of degraded performance in AOD. Specifically, the NMD and NRMSD values for AOD in this experiment appear to have increased significantly compared to those in Figure 2. However, the actual NMD and NRMSD values for AOD in Figure 8 are not explicitly reported, making it difficult to assess the trade-off between surface and column performance. I recommend that the authors clearly report the AOD NMD and NRMSD values in Figure 8 to support a more complete and transparent evaluation of model performance.
17. Figure A5: I noticed that the base run shows the best agreement in terms of the regression slope for PM2.5 dust against SPARTAN observations, with a value closest to 1.0. This indicates strong consistency in magnitude between the model and observations, which is particularly important for evaluating exposure-relevant quantities such as surface PM2.5 dust. While other runs may show improvements in NMD or NRMSD, I would encourage the authors to clarify which performance metric they consider most important in this context, and why. A more balanced discussion of the trade-offs between slope, bias, and correlation would strengthen the interpretation of model performance.
18. Figure A6-7: I am unclear about the meaning of “Emis*” as referenced in the figure caption. Could the authors clarify what this experiment represents and how it differs from the base simulation?
19. Table 3: While the Emis* and subsequent sensitivity experiments reduce the NMD and slope for SPARTAN PM2.5 dust, these changes are accompanied by increased NMD for AERONET AOD. This suggests a trade-off between surface and column performance, which should be explicitly acknowledged and discussed. Deep Blue AOD performance is unchanged: Across all simulations, the correlation (r), slope, and NMD versus Deep Blue AOD remain fixed (e.g., r = 1.00, slope = 1.00 for many runs), which is unexpected. Please priove a brief explanation of why these metrics remain unchanged. It would be helpful to indicate the number of observation sites (N) used in the calculation of these statistics for each metric, especially since previous figures suggested limited spatial coverage (e.g., <20 sites).
20. L539-544: This paragraph, along with the corresponding figures in the Appendix, is difficult to follow. It is unclear why the authors conducted the C48 runs, what specific purpose they serve, and which figures are based on C48 versus C49 configurations. The motivation, methodological details, and relevance of these results are not clearly explained, making it challenging for readers to understand their significance. I strongly recommend that the authors either substantially revise this section to clearly articulate the purpose, setup, and interpretation of the C48 experiments, or consider removing it altogether if it is not essential to the main conclusions.
21. Figure A9-12:The validation of seasonal mean values is based on a very limited number of observational sites—fewer than 20, and in some cases fewer than 10. This small sample size raises concerns about the statistical robustness and representativeness of the calculated performance metrics (e.g., R², NMD, NRMSD). With such limited spatial coverage, the results may be highly sensitive to outliers or regional biases, and may not reliably reflect the model’s performance at larger scales. I recommend that the authors either supplement the evaluation with additional datasets or explicitly discuss the limitations associated with the small sample size.