The authors propose a machine learning based retrieval of XCO2 for an OCO-2 like sensor using a neural network. The overall research direction is important for our field and a promising avenue to deal with the ever-higher data rates of future space borne instruments for trace gas retrievals. However, the study present requires major revisions to be considered for AMT. Most importantly I am concerned that the model might get the right result for the wrong reasons and uses some of the parameters like solar zenith and azimuth angle to estimate the location of a given OCO-2 observation, rather than using the information contained in the measured spectrum.  In essence, the model might simply interpolate XCO2 spatially and temporally instead of retrieving it from the observations.

General Comments:

• Please make explicit what the innovation of this work is compared to state of the art.
• I am concerned that the model might get the right solution for the wrong reason and uses some of the parameters like solar zenith and azimuth angle to estimate the location. Please probe for that. E.g. remove spectral information and repeat your model training. How much does your RMSE increase? Remove spatial information (sun-satellite geometry) and repeat your experiment. Use XAI methods to look at feature importance, etc.
• How would we get an uncertainty estimate from your approach? How do you know when your model fails. 
• What is your RMSE compared to TCCON when you apply your final model to OCO-2 data. How does that compare to the current operational retrieval of OCO-2?
• What do you see as important future work items?
• How would you resolve any step changes from two adjacent retrieval models if you would train one model for each region (as you suggested)?

Specific Comments:

4: not clear what ‘retrieval efficiency’ means

4: ‘enhancing the retrieval’ could mean many things. Please be more concrete.

7: retrieval -> retrievals

8: inversion of traditional optimization models -> inversion of the operational optimization model

11: specify ‘the observed region’

13: ‘high precision’. We care about accuracy not precision. Don’t agree with 2ppm being ‘high’

38: remote sensing is not limited spatially or temporally:  can’t observe at night!

39: vital for future greenhouse gas

46: ‘Enhancing satellite sensor performance alone cannot produce datasets sufficient for monitoring carbon sources and sinks’: Why?

47: Need to explain why ‘efficiency’ is important

53: full physics model

54: atmospheric-surface -> atmospheric and surface

58: why is ‘cost function’ in quotes?

61: optimizations requires

63: how much time do the two processes in the retrieval take up? Are they equally time consuming? How much time do they take up compared to any other calculations that are part of the retrieval?

65: ‘While rigorous, standard nonlinear optimization retrievals lack the speed and scalability required for high-precision satellite-based greenhouse gas mapping.’ That is not true. We have OCO-2, GOSAT, …

74: ‘GOSAT instrument mode,’ : what does ‘mode’ mean here?

76: please expand on the work by Zhao et al (2022). What was their approach, what accuracy do they get.

84: ‘In the present paper, a proof-of-concept study demonstrates a novel machine learning strategy ‘: How is this work new compared to the literature?

90: What does ‘multiple sources’ refer to?

96: why is the study limited to East Asia?

109: Could that lead to step changes at the boarders of two models? If yes, how would you deal with those? What limits you to use one model for all the data? Have you tried that? If Yes, what were the results?

Figure 1: Remove: ‘The map was plotted …’

Figure 2: Add more info to figure description

Would remove Table 2 and just provide source.

128: If you normalize each spectrum why do you need the sun-earth-distance as a feature?

138: Did not understand the following sentence: ‘Although the key retrieval information for surface pressure comes from the O2-A band, machine learning models based on simulated data essentially predict XCO2 by fitting the “correct solutions."’

147: ‘The other angles are provided in radians.’ Does this mean the SZA is given as the cosine? Please make explicit if that is the case.

148: ‘velocity of the satellite relative to the Earth’s surface are input into the model.’: Should that number not be constant and therefore provide not information?

163: Why use this subset of months?

173: this in-sample data -> these in-sample data

175: ‘depicts out-of-sample test results on 5% of the training data that was excluded from model fitting’. Training data that was not used for training is not training data.

179: ‘This discrepancy indicates the MLP-XCO2 model fails to fully capture the underlying upward trend in atmospheric CO2.’ Please run an additional experiment to confirm your hypothesis. Hold out 2018 and use 2016,2017,2019,2020 for training.

Figure 3: Why do you show +- 1% and not any other value?

183: Please add more figures and analysis and show how the biases look like spatially and temporally.

207: Remove sentence: ‘These spectra are detected by the OCO-2 satellite detectors after downwelling absorption, surface reflection, and upwelling re-absorption in the atmosphere.’

209: you already introduced abbreviation for WCO2 etc. before.

243: why is an 1% error impressive?

Figure 5: How did you choose the spectra?

Figure 5: ‘proposed’ forward calculation model: that sounds like you are proposing that model in this paper

266: ‘uncertainty’-> variability

269: repetition of text in line 262

275: need more details on how the CO2 profiles are generated

Figure 8: add similar plot for XCO2

287: Why did you restrict yourself to 2016 data?

Code availability: Please upload the code to train the model, the trained model, and the training/testing data to a public repository.

Author contributions: FX developed the forward model : The Refractor model was already developed

Xie et al. present an ML approach for performing XCO2 retrievals based on OCO-2 measurements. Their work follows the publications of David and Bréon who have before shown the general success of this type of method, applied to the same instrument. This work of Xie et al. implements a similar technique. The added novelty then seems to be the training on simulated data (via the ACOS forward model) that is covering a wider range of XCO2, in order to mitigate the issue of the MLP not being able to follow the growth of atmospheric CO2. There are some more general issues that the authors did not mention, such as quality assessment - how does a science data user know an ML-based retrieval is considered "good" and should be used in a study (or what the uncertainty on the estimate is)? Should people use the ACOS averaging kernels, since the ML-based method does not provide any?

The manuscript is very well written and contains useful figures for the most part. In some places, minor re-wording or additional explanation would be appropriate and helpful, I have listed those below. My major comments would be the following:

1. In Bréon et al. 2022, it is revealed that their ML approach inadvertently resulted in the NN "using" the weak CO2 band as a proxy for geographical location and time. They thus removed the weak CO2 band from the training process. However, the authors of this manuscript do indeed use the weak CO2 band and have not explained as to how they overcome this issue. That would be important information, especially since they are following the general layout of Bréon et al. 2022. Maybe the issue does not manifest itself due to the much smaller region of interest, but the authors must show that.
2. Related to (1), the "glitch" discovered in Bréon et al. 2022 was only found after they investigated specific features present in the original ACOS OCO-2 retrievals which were missing in the ML-based retrievals (specifically, strong plumes). The authors here only look at broad bulk-type statistics by comparing to TCCON and to ACOS OCO-2 via simple scatter plots. The strengths and weaknesses of the proposed ML-approach should be investigated more thoroughly by analyzing the results more carefully. Do the same biases appear in the derived retrievals, compared to the training data? Are global-scale features retained just like regional and small-scale ones? Are new biases introduced? There is possibly more to learn from the data than is shown in Figures 10 and 11. While the approach is promising, the authors should attempt to show an assessment of the quality of the ML-retrievals beyond the simple scatter plots.

My suggestion to the authors would be to (1) demonstrate that their approach, while using the weak CO2 band, does not result in a loss of local features, such as plumes (analogous to Bréon et al. 2022, Figure 4). Further (2) they should demonstrate that their ML-based retrievals retains other characteristics of the training set (regional-scale, or local-scale; observe differences on maps etc.)

Minor suggestions:

Line 4: "low retrieval efficiency" and "insufficient retrieval accuracy" are somewhat diffuse terms; I would simply mention challenges regarding computational efficiency.

Line 38: "not limited spatially or temporally" is not quite true - space-based platforms have observational coverage in space and time as a result of their orbital characteristics and other instrument parameters.

Line 62: The interpolation of absorption coefficients for the calculation of optical property inputs for RT calculations are generally quite fast and can be done in less than a second typically, if the code is optimized enough (amongst other things). The computational effort is mostly driven by the RT calculations.

Figure 5: It is not fully clear to me what these represent. Did the authors take the outputs of the ACOS retrieval L2STD products and use them as input in their ReFRACtor-driven set-up? Please clarify.

Line 338: When discussing the computational effort, it is mentioned that the forward model takes 12.16s to process two bands; but in forward-model "mode", Jacobians are presumably not calculated, so the actual retrieval set-up would be even slower than the mentioned 36.48s. It would also be very interesting to learn how long the training process took!