Review of the manuscript

Air quality modeling intercomparison and multi-scale ensemble chain for Latin America

Pachon et al., 2024

The submitted manuscript describes an intercomparison of global and regional air quality models operating over Latin America, focusing on the model performance over 4 selected cities. The model results are evaluated against surface measurements of the main air pollutants (NO2, O3, SO2, CO, PM), for each individual model as well as for a model ensemble. The paper in detail describes performance of each model and of the ensemble median by providing the statistical scores for selected pollutants, cities (of different sizes) and seasons (January and July). The study raised up number of interesting points, such as the need for higher model resolution over smaller cities, importance of emission inventories including the local knowledge, the role of wildfires, etc. It also addressed the capabilities and limitations of each of the model and its setup.

As the authors state, this is the first study focusing on the intercomparison and evaluation of the air quality models in this region, and I find this coordinated effort a unique and valuable step forward for the air quality modeling over Latin America.

I recommend the manuscript to be accepted for publication after addressing the following minor comments:

1.   Since the paper compares results of different models and their set-ups, I find the Section 2.1 (Description of the models and modeling set-up) to be the core part of the manuscript. However, it seems a bit inaccurate or lacking important details. Could the authors please be more specific and include in the paragraph describing each model information on the meteorology driving the model, anthropogenic, biomass burning and biogenic emission inventories, stating the exact name of the emission datasets?

E.g. for SILAM the authors say “the anthropogenic emissions were adopted from the CAMS global emission inventory” (L 74). However, there exist different CAMS inventories and different versions. Also, when the authors say “The biogenic emissions were simulated off-line by the MEGANv2.1 model” (L68, L76, L83), does it mean the MEGAN model runs were performed specifically for this study or did the model use an offline emission inventory calculated by the MEGAN model? Please make clear and if the latter, please specify which biogenic emission datasets were used.

The paragraph describing the WRF-Chem set ups (both for MPIM and USP) is rather brief (L97 – 100). Could the authors be more specific and provide more detail on the emission data used and the difference between MPIM and USP set ups?

The above mentioned applies also to the summary Table 1. The descriptions seem inaccurate or incomplete.  

• The table is missing vertical resolution for MPIM WRF-Chem and projection for ECMWF-CAMS – please add and if not possible to define, indicate so in the table
• Please define IC-BC abbreviations in the text or in the table footnote
• Please be more specific in description of the emission datasets used and state the name of the emission dataset (including version). E.g. for SILAM the Table states CAMS-REG-AP v3.1 and TNO-MACC which are both regional European inventories. But it is not clear which global anthropogenic dataset was used. Similar for biogenic dataset.

2.      South America, esp. the Amazon, is one of the major sources of biogenic VOC emissions globally. I would expect the biogenic VOCs could impact O3 and CO concentrations, esp. in Bogota and Sao Paulo. The paper discusses effect of NO2 on O3, mentions effect of wildfires or excessive OH concentration on CO. But does not mention the possible role of BVOCs. Could the authors please comment on this and where appropriate, include the effect of BVOCs in the discussion? E.g. could the model underestimation of CO be partly explained by possible underestimation of BVOC emissions? E.g. The CO January maxima “north of Argentina, south Bolivia, Paraguay and south of Brazil” (L433, Fig. 7) coincide with locations with high isoprene emissions.

Technical comments:

I’d suggest adding a short paragraph at the end of the Introduction section, overviewing the following sections of the manuscript.

L90: please remove scales (repetition)   

L91: please replace FINN module by FINN dataset

L115: please replace Suplhur by suplhur

L116: please add PM10 as well

L133: Please replace simulate by simulated.

L242: The sentence beginning ‘On the other hand’ seems incomplete.

L395: Please check the MNBIAS and FGE values in the text. According to the Table A4 these should be 3.6% and 0.1.

L415: Please replace ‘hot pollution spots’ by ‘pollution hot spots’

Summary:

* This paper does important work comparing simulations over lesser studied regions.

* The paper is written as though motivated by forecasting, but the methods seem more focused on historical application and does not provide much discussion of forecasting needs/limitations.

* The paper has many endpoints and many locations. The current discussion that starts with individual species and all locations was somewhat difficult to read. It would be nice to provide high-level context and specific useful details.

* Overall, the paper has an impressive scope, but could improve readability/organization/focus.

* Though I recommend larger improvements, the article as is represents a substantial effort and could be published with minor updates.

High level thoughts:

* 2015 seems like an odd choice for a precursor to a forecast system. Perhaps frame it a bit differently.

* Is city size really the determinant factor? Or do Bogota and Santiago have other challenging features for models? While this is interesting, it should be frame as a correlative speculation.

* Abstract: what was the suite of endpoints? You say O3+NO2: good; SO2: bad; ?: moderate.

* Mediation of outliers doesn't seem like "demonstrating the potential to establish an analysis and forecast system". Was it tractable? Could a single model have performed similarly? If this is setting the stage fore future application, I'd like to see more discussion of the application-specific pros/cons.

* Did this system out-perform publicly available forecasts from CAMS and GEOS-CF? Since they did not report in 2015, it is hard to say... However, if existing global forecasts are outperforming your ensemble, the goal of forecasting is already solved. I think it would be good to set the stage for the need for a local forecast.

* Both global models are based on C-IFS meteorology and several regional models too. How does this influence the spread of the ensemble?

* I suggest including lots of statistics (as you did), but only in the appendix. In the text, the paper would be improved by focusing on a few. Given that you use city averages, you have just 30 points of data so presenting so many statistics seems disproportionate to the data populating them.

* Similarly, the per species sections with obs, model, inter-model makes the paper quite long for the value. If the median ensemble was added with the individual models you could reduce the paper length.

* The comparison of model area mean to monitor mean is not a particularly useful comparison. Monitoring networks are typically located in a spatially biased manner. They tend to be located near high concentrations and near people. As a result, when comparing the observation to the model, only the model is a spatial average. The observation is spatially weighted. A more fair comparison would be to sample the model at observations and then average it. In that way, the model would be spatially weighted in the same way as the observations. Or, you should attempt to remove the monitor location bias. You could do that by averaging the monitors within pixel/polygon intersection. They perform the same weighted average on the observations. In short, right now you are applying meaningfully different spatial averaging on the model and monitors. The more complex method does not seem to address this.

Dataset Methods:

* CAMS emissions are described in detail by specific reports. You should consider citing those documents rather than MEGAN generally.

  * https://eccad.aeris-data.fr/essd-surf-emis-cams-bio/

  * https://eccad.aeris-data.fr/essd-surf-emis-cams-ant/

  * https://eccad.aeris-data.fr/essd-surf-emis-cams-soil/

* Table 1:

  * Global models do have initial conditions (IC), so "global model" is not sufficient. How long were they spun up? From what?

  * Vertical structure is important, but perhaps most relevant is the depth of the first layer which directly influences the comparability of the model to the measurement (at a few meters). Also, I recommend being consistent. 25 layers up to what? 60 levels up to what? 35 levels up to what? MPIM vertical structure?

  * Recommend adding cell size to the projection cell. Because distortion varies by projection type, this would be useful to understand where and when ~0.2 degree is achieved.

Large versus small areas:

* The discussion of city-wide means would benefit from intra-urban variability of model performance. Are large cities (e.g., Mexico) simply averaging high and low biases in the city-wide mean? The large cities also have complex topography/coastal issues that could lead to problems from 0.2 degree resolved models.

* You likely have sufficient observations within Mexico and Sao Paola to say something more than simply observing a difference in means.

Specific questions:

line 118, "modified normalized bias" should probably be "mean normalized bias" as described in Table A1.

line 126, 75% completeness for the entire dataset? Was there any completeness applied to specific days (e.g., 18 of 24h)? Was there any minimum number of sites per city? Per species?

line 128-129, if all the data is used, then the appendix should include data completeness and number of monitors for all cities. I do not see that.

line 139, "In all cities the data availability was 100%." It sounds like you're saying that all stations in all 8 cities never missed a single hour of measurement for the simulation period. That sounds amazing, but I could be misinterpreting the sentence.

Figure 2 shows considerable missing data from CHIM. I didn't see anything about models being incomplete, so this seems odd. How does this affect the interpretation of statistics from CHIM to other models? After reading the appendices, I believe that missing data is described there.

Figure 2 shows that no model has a prediction on July 31, but two models in the appendix claim 100% data coverage. One of these two things is not true.

line 147, [the] NO2 [mean] is underestimated [by all ensemble members]

line 149, what was the Bogota value?

line 149, "the model fields are above and below" do you mean that the ensemble members both over and underpredict?

line 165, does this suggest the correlation is related to meteorology and the magnitude to emissions?

line 186, Please rewrite this sentence.

Table A1: I wonder if Coefficient of Variation has a typo. I am used to this being the std dev divided by the mean, which given a constant mean is larger when the variance is a larger. Because your table does not report standard deviation and the description does not either, I can't check what the results are.

line 189-191, does dispersion mean the variation between ensemble members? I'd recommend not using the word dispersion because air quality scientists use this word to describe ventilation.

line 194, the number of sites with obs should be added to figure 1 using different shapes for ozone, no2, co, so2, and pm. It seems odd that a reader would have to go to the appendix to know. In the text, at least provide a range for the four cities that are the focus of the paper. Is it between 1 and 10?

line 196-197, ozone season for Mexico? Are all the areas the same?

Figure 3, I recommend using a common y-minimum (e.g., 0 ppb)

line 207, is that high or low for ozone in Bogota? I shouldn't have to check the figure to understand the text and vice versa.

line 386, which one?