The authors construct a machine-learning emulator to generate atmospheric transport simulations for greenhouse gas and air quality modeling. Atmospheric transport simulations are a major computational bottleneck in greenhouse gas and air quality modeling, and I think this paper addresses and important topic. I enjoyed reading his paper, and I think it'll make a nice contribution to the literature. I have several suggestions for editing the paper:

High-level suggestions:

I think the article would benefit from much more extensive evaluation of the emulated footprints. The figures in the manuscript examine the correlation between the log of the true footprints against the log of the emulated footprints. At the end of the day, many modelers ultimately care about the accuracy of simulated greenhouse gas or air pollution mixing ratios. Hence, I personally think that the correlation between the footprint values is necessary but not sufficient to convince many modelers (including myself) to use a tool like FootNet. For example, let's suppose one used these footprints to model CO2 and CH4 mixing ratios. Would these footprints capture peaks and troughs in CO2 and CH4 (I.e., suppose one were to plot CO2 and CH4 at individual sites as timeseries.)? Would the emulated footprints capture spatial variability in atmospheric CO2 or CH4 levels? Suppose there were CH4 super-emitters scattered in the Barnett Shale. Would the emulated footprints accurately capture the impact of those super-emitters on downwind atmospheric observations? Let's say one were to model CO2 using these emulated footprints. Would those footprints capture diurnal variability in CO2 mixing ratios (i.e., due to variability in both fluxes and boundary layer dynamics)?

In addition, I imagine it might not always be practical to use 85% of data for training. For example, if one wanted to run footprints for a large satellite dataset, it might (hypothetically) only be feasible to use 5% or 10% of the data for training. In that case, one would need to train FootNet on a limited number of data points and then run the trained FootNet algorithm on a much larger number of data points. Do you have a sense of how FootNet would perform in this circumstance? Both of the cases studies described in this paper are for small geographic regions (e.g., San Francisco and the Barnett Shale). Let's say one wanted to use FootNet across the entire US or across the entire globe. For these larger spatial scales, I imagine there are more variable and diverse transport patterns in different regions. In this circumstance, one would want FootNet to capture all those transport patterns in different regions. By contrast, San Francisco and the Barnett Shale, by factor of their limited geographic size, might have a more limited set of transport patterns to capture. These different circumstances might necessitate very different approaches to the training data, and the resulting emulated footprints might not have the same fidelity or accuracy.

Specific suggestions:

• It would be helpful to include line numbers on future versions of the manuscript. Doing so would make it easier to discuss specific lines of the manuscript.
• Abstract: What does "near-real-time" mean in this context?
• Intro: "The sensitivity of each receptor to its upwind sources, termed as the receptor’s “footprint”, can then be used to estimate fluxes inversely (e.g., Turner et al., 2020)." There are a bunch of other good references that could be used as examples here going back to the early 2000s.
• Intro: "The footprints are integrated 72 hours backwards from the measurement time." A lot of continental-scale studies using STILT use footprints that are integrated 10 days backward from the measurement site. Do you think the approach developed here is applicable to those longer time scales?
• Sect. 2: "Each convolutional block includes two convolutional layers with 3 × 3 convolutional kernels and one 2 × 2 max-pooling layer." What is a convolutional block, convolutional kernel, and max-pooling layer? I suspect that most readers won't be familiar with these terms.
• Sect. 3 "The overall correlation between FootNet predictions and STILT simulations..." Does this line refer to r or r^2?
• Pg. 8, line 1: This paragraph feels like it could use a better topic sentence. You've just finished describing the training process for a footprint calculation in the Barnett Shale. What topic or concept are you going to describe next? I think the answer to this question would better guide the reader and give the reader a better idea of what to anticipate.
• Fig. 3F: I was a little confused about Fig. 3F. Are the individual footprints summed before being plotted in this figure? I.e., does this figure compare the log sum of each predicted footprint against the log sum of each true footprint? Alternately, is each individual model grid box from each footprint a different point on this plot? I imagine that the former plot would show less noise and a higher correlation coefficient whereas the latter plot would show more noise and a lower correlation coefficient. I would recommend clarifying how this figure is constructed.