General comments

This is an interesting application of Machine Learning techniques to produce a stratospheric aerosol extinction profile data set based on limb-scatter measurements with OMPS-LP instruments. The approach seems to be very promising and the manuscript is suitable for AMT. The manuscript should eventually be published in my opinion, but I ask the authors to consider the following comments.

I have one general comment regarding the role of O3. The current approach does not consider O3 in any way. However, O3 can significantly affect the limb-radiance profiles in the visible spectral range. I assume that neglecting O3 is one main reason, why the performance of your approach is poorer at high southern latitudes. I also assume that the differences between your NRT data product and the V2.1/V2.2 data product are larger during the O3 hole season than in the temporal average shown in Fig. 1. Please also show the temporal variation of the differences for selected latitudes to investigate this further.

Just to make this clear: I’m not asking you to include O3 in your approach at this stage. I think the paper is worth publishing without an explicit consideration of O3. But consideration of O3 may significantly improve the performance of the NRT aerosol extinction retrievals and may be done in the future.

Specific comments

Line 1: „Among aerosol characterization methods“

This is only a minor point, but why "characterization"? Does “characterization” include simulations and observations? I've never seen "characterization" in this context and its meaning should perhaps be explained?

Line 2: “characterization”

Line 3: “but the radiative transfer modeling of scattering processes performed by traditional retrieval methods are too computationally expensive for near-real-time applications without simplifying assumptions.“

I’m not sure if this is really true. It will depend on the instrument, the number of limb measurements per orbit, the computational resources available. I’m aware of algorithms that can process an orbit in, e.g. an hour, making NRT retrievals possible, of course depending on the specific meaning of “NRT”.

Line 16: “cooler temperatures” -> “lower temperatures” ?

Line 35: “The Ozone Mapping and Profiler Suite (OMPS) Limb …”

This sentence is incomplete.

Line 41: “Gauss-Siedel” -> “Gauss-Seidel”

Line 46: „their runtime is prohibitive to NRT applications unless compromises are made”

How long does it take to process one orbit with the method of Taha et al. (2022)?

Line 52: “that produced that connect inputs and outputs”

Something seems to be missing/wrong here?

Table 1, caption: “Date ranges of OMPS LP data considered.”

Considered for training, validating, testing?

Line 67: “We utilize measurements during specific periods between October 2013 and December 2022 (Table 1).”

It is unclear, whether these periods are used for training the NNs? Or testing?

Line 76: “calculated from quantities available in L1G”

Please explain “L1G”

Lines 71 – 78: what about O3? Isn’t it necessary to consider O3 in some way? The Chappuis bands will have a significant effect on the shape of the LR profiles in the visible part of the spectrum.

Line 80: “These inputs correspond to the aerosol extinction coefficient reported in the OMPS LP aerosol retrieval version 2.1 data product”

Context is unclear? What does "These inputs" refer to? The extinction coefficients should be outputs (of the NN), right?

Line 83: “we assume a value of 10^-8”

Unit is missing (1/km)

Line 85: “NN to be less than 10^-8 are replaced with a fill value of -999”

Again, the unit is missing. What happens if the predicted value is e.g. 1.1 x 10^-8?

Line 92: “These correction methods introduce differences in the retrieved aerosol extinction coefficient, and so our input-output pairs do not have a perfect one-to-one relation.”

This does not seem to be ideal, because the input/output data sets used for the training are not consistent. Can you quantify the effect on the estimated/retrieved aerosol extinction coefficients?

Line 99: “These data are split into training, validation, and test sets.”

I’m not sure, how this relates to the periods listed in Table 1? Are you splitting the periods listed in Table 1? In the caption of table 1 you mention that 10% of the data are selected (for what purpose?). I’m not sure, how this fits to the 70%, 20%, 10% splitting mentioned in line 101. Please remove the inconsistencies.

Line 122: “as we found poor performance at 510, 600, and 675 nm in the southern hemisphere,”

I’m not too surprised by this finding, because O3 will affect this spectral region and the SH experiences very low O3  concentrations at higher latitudes during the O3 hole season.

Line 146: “The notable exceptions to this are lower altitudes in the southern hemisphere at the shorter wavelengths”

See previous comment.

Figure 1: It would be interesting to see the differences in concentrations as a function of time, particularly in the SH at high latitudes. I assume that during the O3 hole season the differences can be significantly larger than in the temporal mean. Please show a time-altitude-contour plot of aerosol extinction differences for different latitudes.

Caption Fig. 1: “The dashed line is the tropopause altitude”

Where does the tropopause altitude data come from?

Fig. 2 and Fig. 3: “Le Soufriere” -> “La Soufriere”

Fig. 2: Please explain „WC“

Fig. 2: How is the sAOD determined, i.e. what altitude range/tropopause data?

Fig. 4: "Tangent height" should read "altitude" or "height", right?

Fig. 4: Unit is missing (1/km)

Caption Fig. 3: “Ha’api” -> “Ha’apai”

Line 185: “hardware, achieving a ~60x speedup compared to V2.1.”

OK, so a full orbit takes about 120 minutes with the V2.1 processor, compared to 90 minutes orbit duration. This means that with some adjustments it would be possible to process one orbit in 90 minutes with the "full physics" version.  Also: what machine are you using for the calculations?

Line 190: „When experimenting with changes to the radiative transfer-based aerosol retrieval algorithm, our methodology can therefore significantly reduce the computational resources required to determine how such changes would affect the mission’s complete record.”

I'm not sure this would really be the case. If you have differences of 20% and more between the NRT and the full physics data set, how would that help to test how changes would affect the entire record? Perhaps I'm missing the point here?

Fig. 5: "Tangent height" should read "altitude" or "height", and the unit is missing (1/km).

Fig. 6/video supplement: The unit (1/km) is also missing in the video supplement.

Line 195: “with strong biases in the southern hemisphere and shorter wavelengths”

Did this occur in all seasons? Again, I presume that O3 and the O3 hole play an important role here.

Line 196: “as well as OMPS LP’s Sun-synchronous orbit”

You mean the large scattering angles in the SH and the small ones in the NH? This is not only a consequence of the sun-synch orbit, but also of the viewing direction. OMPS-LP could also be viewing in the opposite direction. Then you would have small/large scattering angles in the SH/NH.

Line 207: “This confirms our approach’s implicit assumption that the NNs can learn to handle the minor differences in corrections applied to the radiances between versions 2.5 and 2.6,”

I'm not sure that this implies that the NN can handle the differences between versions 2.5 and 2.6. That would be surprising, right? It probably means that the differences were not so large!?

But it would be very interesting to mention the differences in performances in a quantitative way.

Line 216: “In the former case”

Not entirely clear what “former” refers to? High altitudes or SH? or both?

Fig. 7: "Tangent height" should read "altitude" or "height", and the unit is missing (1/km).

Fig. 8: "Tangent height" should read "altitude" or "height", and the unit is missing (1/km).

Conclusions: perhaps one could mention as an outlook that the approach could be extended to consider O3 as well? This would probably improve the performance of the NRT data product.

Line 280: “that the NNs properly” -> “that the NNs are properly”?

Review of “Using neural networks for near-real-time aerosol retrievals from OMPS Limb Profiler measurements”

This paper presents a method of retrieving aerosols from the OMPS-LP measurements using neural networks trained on the current NASA v2.1 aerosol product. The proposed method is much faster than the current v2.1 approach and provides similar results under both background and elevated conditions. The retrieval proposed here could be used for both NRT applications as well as testing of future algorithm updates, via training on a limited set of test data. The paper is well written, with results presented clearly and convincingly. Given the broad interest in machine learning applications, potential computational improvements and relevance to NOAA-21, I would recommend publication with minor corrections.

General Comments:

As someone without experience in neural networks, I found the description in section 2.2 somewhat difficult to follow. This is probably standard in ML literature, but given the atmospheric journal/audience I think some context would be helpful here. For example, values for the various activation functions, epochs, batch sizes, cyclical and minimum learning rates are nicely provided, but what these terms mean and how these choices impact the retrieval are not clear except for a brief mention of overfitting. I don’t mean to turn this into a primer on ML, but a bit more of a link between these parameters and the results would be nice.

A lot of emphasis is given to the computational cost of v2.1 and the inability to produce a near real time product using this approach. However, I don’t think these conclusions are justified by the paper in its current form. For example:

Line 4: “processes performed by traditional retrieval methods are too computationally expensive for near-real-time applications without simplifying assumptions.”

OMPS retrievals as implemented in v2.1 are an “embarrassingly parallel” problem across profiles, slits and wavelengths and seemingly could be sped up if needed. This may be cost prohibitive, but the paper does not discuss what hardware is currently required to process v2.1 or potential increases needed for a real time v2.1 algorithm vs NN. As it is, the OMPS-LP v2.1 data is processed at one day per day, so it seems the throughput of the current v2.1 system is not a problem, only potentially the lag.

Similarly, the cost of running the v2.1 algorithm and NN algorithms is discussed (line 185), but it's unclear what hardware was used for the comparison. Is the NRT approach ran on the same hardware as the v2.1 retrieval? Or are there special hardware requirements for this NRT retrieval (GPUs?) It is mentioned that 97% of the time is spent loading the NN into memory, does this require a machine with a large amount of memory (or more than the v2.1 uses)? Is the training factored into this analysis of computational cost?

Specific Comments:

Line 46-48: Is the time to process the retrieval of a single profile in v2.1 so long that it precludes NRT applications? I would have guessed (maybe incorrectly) the downlink, attitude solution, L1 calibration, atmospheric reanalysis etc. would have been a larger contributor to any lag in NRT products than the retrieval itself. How “NRT” could the NN version be in practice, given this is proposed as a major benefit of the proposed system?

Line 71: What are the outputs of these input-output pairs? Is it cloud top altitude, enhanced layer and PSC, as marked in Figure 4 as well as multi-wavelength extinction?

Line 94: I would change to “no NASA OMPS LP aerosol retrieval version” as the University of Bremen and University of Saskatchewan OMPS-LP aerosol products both use version 2.6.

Line 95: What is meant by “differences in correction methods are consistent”?

Line 125-130: Probably obvious for someone in the ML field, but how are results from these two NNs put together?