This manuscript analyzes the two-decadal (2003-2024) AOD variability and trends using the AOD from the 4th generation of the ECMWF atmospheric composition reanalysis (EAC4) that assimilates the AATSR and MODIS satellite AOD into the CAMS model. First, AOD from EAC4 and CTRL run without assimilation are evaluated with AERONET observations from 178 stations using seven specific metrics (R, MAE, RMSE, RMB, %MFE, %FGE, and IOA). The evaluation shows better agreement between EAC4 and AERONET than between CTRL and AERONET. Second, the global and 18 regional AOD trends over the two decades (2003-2024) as well as over the first (2003-2013) and second (2014-2024) half are examined for both EAC4 and CTRL to understand the impact of satellite data assimilation. Third, the AOD trends from individual aerosol species of dust, sulfate, organic matter, black carbon, and sea salt from EAC4 in 18 regions are assessed to gain insights in explaining the contributions of aerosol species to the total AOD trends. 

I found this study is thorough and robust. It provides quantitative evaluations of the quality of AOD reanalysis and model simulations with multiple evaluation metrics and reports the 22-year AOD trends in different regions of the world. The results are appropriate for publications on AMT/ACP. However, it is a very long paper that takes quite a bit effort to go through. I would suggest either shorten it to make it more concise or break it into two papers, if the authors are willing. In addition, I do have several major comments and many specific ones that should be addressed or clarified before publication.

Major comments:

1. AOD assimilation: 
   
   (1)    It is described that in 2003-March 2012 the AOD from both AATSR and Aqua-MODIS were used to produce the EAC4 AOD reanalysis, but after that only Aqua-MODIS was used. How consistent is EAC4 AOD in the two periods (before and after March 2012)? Have you examined the differences by excluding AATSR in data assimilation for 2004-March 2012 to quantify the effects?
   
   (2)    It is not clear to me how the fractions of speciate AOD get adjusted during assimilation (e.g., the pie charts for speciated AOD fractions from CTRL and EAC4 shown in Fig. 4b and Fig. 6). Since there is no information about aerosol composition in satellite AOD data, the speciated AOD fractions are expected to be very similar between CTRL and EAC4 even though the values of total AOD are changed after assimilation.
   
   (3)    Related, it is important to clarify that speciated AOD, or AOD type as referred in the manuscript, are not reanalysis products and should not be misused.
2. Missing components in CAMS AOD:
   
   (1)    Nitrate: The CAMS aerosol omits one of the major aerosol species – nitrate. Numerous observations and modeling studies have shown that nitrate aerosol have emerged to be one of the major aerosol species especially in urban and agriculture environments where its amount is similar to, or even higher than, sulfate aerosol. I understand the complexity in incorporating nitrate aerosol in global models because chemical productions from multiple precursor gases and the thermodynamic equilibrium processes must be considered. However, the errors and implications by omitting nitrate aerosol should be at least acknowledged, estimated, and discussed, especially when the speciated AOD trends are explained.
   
   (2)    Volcanic aerosol: It is not described in the manuscript how volcanic aerosols are dealt with in the CAMS model, such as emissions of precursor gases, plume injection height, and timing for sporadically erupting and constant degassing volcanic emissions.
   
   (3)    Stratospheric aerosol: It seems, from previous publications (e.g., Benedetti et al., 2009) that the aerosols in the stratosphere are not included in the model. Although in general stratospheric aerosols are a small fraction of total AOD except in cases of large volcanic eruptions and pyroCb fires, omission of stratospheric aerosols should be acknowledged and discussed for attributing contributions of individual aerosol types.
3. Impact of satellite data assimilation on total AOD:
   
   It is nice to see the comparisons of AOD evaluations, trends, and variability for both EAC4 and CTRL in this manuscript, and the differences between EAC4 and CTRL should be solely attributed to the satellite AOD data assimilation. However, it is not elaborated what can be learned from the differences between EAC4 and CTRL in terms of physical processes in the underlying model improvements in addition to the possible over- or under-estimation of some emission sources. Afterall, data assimilation can bring the results closer to the observations without fully understanding of the drivers for the model-data differences in the process level, but it is the model that carries the burden for understanding what happened in the past and projecting the future change. 
4. Evaluation parameters: 
   
   Seven metrics (R, MAE, RMSE, RMB, %MFE, %FGE, and IOA) are introduced to evaluation AOD from EAC4 and CTRL, measuring absolute, relative, directional, the time evolution agreements. Such suite of metrics is comprehensive, but on the other hand it is hard to grasp the overall performance. Is it possible to develop an overall score or reduce the number of metrics without sacrificing the quality of evaluation? In several cases in the text, only a subset of these metrics was quoted to show the agreement between EAC4 (or CTRL) and AERONET anyway.

Specific comments:

• Line 51, AOD “represents the column-integrated attenuation of solar radiation”: this is only applicable for AOD at the short-wave solar spectral range. I would just delete “solar”.
• Line 75: I wonder what is the advantage of including AATSR AOD, or what is the disadvantage without including it.
• Line 76-77: If all models assimilate same satellite AOD, of course the model spread would be reduced compared with free-running simulations. But is this what models should do and for what purpose?
• Line 85, “good agreement” – this is a subjective statement. What is the standard of “good agreement”? In general, such subjective phrases should be avoid but quantitative measures of agreement should be given.
• Line 90-92: Degassing volcanic emission are widespread. Why is the bias only existed at Mauna Loa and Mexico City? Back to my main comments, how is the volcanic aerosol modeled (emission, injection height, chemical formation, and removal)?
• Line 94-95: Again, “performed well” is subjective. What are the values of “high correlation, low error metrics”?
• Line 123: spell out “MIDAS”.
• Line 166: the horizontal resolution was stated at 0.75 deg by 0.75 degree in line 205. 
• Line 170-171, size ranges: Are they in radius or diameter?
• Page 6 description of emissions: what is the volcanic emission dataset used in this study?
• Line 190-191: Is the increased bias due to the loss of AATSR? How does AOD compare between AATSR and Aqua-MODIS in their
• overlapping period? What is the magnitude of this “small increase”?
• Line 201-204: I cannot tell how FRtype would change after assimilating satellite AOD. 
• Line 213-214: “…typically from 340 to 1640 nm, covering both visible and near-infrared regions”: 340 nm is in the UV spectral region.
• Line 221-222: Is the model monthly mean AOD calculated to match the time and location availability of the AERONET AOD?
• Line 227-229: I still don’t understand why the volcanic emission would introduce a challenge for global models over particularly Mauna Loa and Mexico City. Is this a problem specifically in CAMS model? What the injection height database does CAMS use?
• Line 236: “annual mean” for a particular year or 22-year mean?
• Line 275-277, and Eq. 8: Why should the range of FGE be the same as that of MFE at 0-200%? From Eq. 8, FGE could be negative, which is also shown in the calculations presented in this study.
• Line 300-302: Zmk has not been used to indicating the significance of the trend in this study. Instead, p-values are used.
• Line 314: Spell out “ROI” where it first appears.
• Line 353: “good agreement” – again, please be objective and quantitative. How good is good?
• Line 355: FGE is 0.78% on the right panel of Fig. 3.
• Line 358: FGE is -20.71% (negative value) on the left panel of Fig. 3.
• Line 365: what is “(f)”?
• Line 376: Avoid using subjective phrase “good agreement”.
• Line 395: “noticeably narrower spread” – I cannot tell if the spread of delta AOD is narrower in EAC4 than in CTRL from Fig. 4b. Is there a quantitative measure of the spread, e.g., full width at half maximum, or standard deviation?
• Line 402-405: Again, it is not clear to me how the fractions should change after assimilating satellite AOD without information of aerosol composition in satellite AOD.
• Line 427-428: Give numbers of “high correlation…small biases”.
• Line 429-430: Give quantitative numbers of “close to unity”, “a large fraction of stations”, and “large fractional errors”.
• Section 3.1.2: this section is mostly descriptive. I suggest put the statistics in a table in the main text.
• Line 449-450: I disagree that the assimilation consistently improves regional model performance. Assimilation of reliable satellite AOD data does produce more accurate AOD product, but I don’t see any evidence that the model itself is improved. Can assimilation help understand what processes that need improved from the CTRL configuration?
• Line 454: “persistent underestimation of coarse-mode dust”: Not all dust are in the coarse-mode. For AOD at 550 nm, fine mode dust should contribute significantly to the dust AOD.
• Line 461 and 462: Again, “reasonably well” and “exceptionally well” should be avoided.
• Line 464-465, overestimation of AOD remains in EAC4: I wonder why in this case assimilating satellite AOD does not help reduce the bias. Does satellite AOD also show such bias compared to AERONET?
• Line 488: What is the highest AOD values (e.g., AOD > xx)?
• Line 493: Again, not all dust are coarse-mode dust.
• Line 493: What regions are dominated by transported fine Saharan dust?
• Line 496-497, EAC4 appears to underestimate AOD in NAC more strongly than CTRL that is explained by uneven sampling across AERONET AOD bins: I don’t understand – The “uneven sampling” should occur in both EAC4 and CTRL and in other regions.
  
  Line 522-523: Again, I don’t understand how satellite AOD assimilation modifies the fraction of individual aerosol components using Eq. 1. Although AOD changes after assimilation, the relative contribution should remain the same since no information about the composition from satellite AOD.
• Line 683, biomass burning regions: the current NAC is not just dust, but also includes part of important biomass burning area. I wonder why not separate NAC into dust and biomass burning source regions (0-15N)?
• Line 683-685: Not much fire-induced maximum AOD values in June in all biomass burning regions, and very low AOD in Nove over AMZ and SEA, according to Fig. 9.
• Line 685-686 and Fig. 9: the enhanced AOD in May-Sep 2023 is very hard to see in Fig. 9 because the low contrast of colors between the 2023 AOD and background.
• Line 702: Can you exclusively attribute the increasing AOD trend over ME to the rise of anthropogenic emission, considering that more than 1/3 of AOD over ME is dust?
• Line 712, Jan-Feb peak AOD in 2020 is mostly due to the Australian New Year’s fire.
• Line 721: According to Fig. 2, dust contributes ~0% to total AOD in SPO and SAIO.
• Line 724: “…contribute significantly to the observed AOD variability over the oceanic region” – this seems to contradict with the sentence in line 710 “Over the oceanic regions, AOD exhibits a generally constant pattern.”
• Line 747-749: Positive trends not just extend toward Sahel – most area in N Africa show positive trends.
• Line 827-829: Explaining the positive CTRL DOD trends to the additional dust emission from the land use change after biomass-burning: Such process is not considered in the CTRL simulation to explain the DOD increase.
• Line 886: Add references for “reported reductions in biomass-burning emissions in recent years”.
• Line 887-889, declined fire activity because of the increased drought frequency that reduces the fuel availability: it is often found contrary – increased drought reduces the fuel moisture thus cause more fires.
• Line 892: Again, not all dust are in coarse-mode.
• Line 899-901: I don’t see the consistency of timing in DOD trends in Fig. S27 with the total AOD trend.
• Line 915: “very good agreement” – how good is very good?
• Line 983-984, AOD trend over EC: There is a statistically significant strong increase trend in the first decade.
• Line 1042-1043: Looking at Fig. 16, I think for windows beginning 2008, the trends weaken and gradually become more positive with no statistical significance.
• Line 1043 and line 1046: What does “partial reversal” mean?
• Line 1044: I’d say “with almost all windows exhibiting weak positive tendencies that are statistically insignificant”.
• Line 1059: I don’t think the AOD trends over ADTZ is more heterogeneous than that over WME.
• Line1089: Here only 4 out of 7 metrics of evaluation is used. Does it mean that the other 3 are not critical and can be left out?
• Line 1091: The main reason EAC4 outperform CTRL is because MODIS AOD is more accurate than CAMS simulated AOD in the CTRL run, not because of the assimilation technique itself.
• Line 1093: Again, there is no information to attribute the under-representation of dust to either fine or coarse mode dust.
• Line 1098: What is MBE?
• Line 1102-1104: The issues of complex terrain, coarse model resolution, satellite retrieval uncertainties, and emission uncertainties are not unique for Taiwan and northeastern China; why the discrepancies are particularly large over those regions?
• Fig. 7 and Fig. 8: It is very hard to read the trend color especially over the stations in tringles outlined with lime color. All one can see is the lime color. I think the outlines are unnecessary since the stations with statistically significant trends use triangle symbols instead of circles.
• For all sliding window figures (Fig. 13-16, and Fig. S28-S51): It is hard to see the trend color under the lime-colored hatches when the color is light. I suggest remove the hatches but keep the lime-colored outlines for indicating significances.