The study by Johnson et al. applies a novel inversion technique, i.e., the Gaussian Process machine learning method to infer state-wide, source-specific CO₂ fluxes using OCO-2/3 XCO₂ data, illustrating that OCO-2/3 XCO₂ can be assimilated into inverse models to estimate sub-regional and source-specific CO2 fluxes on a seasonal- and annual-scale. In general, this is a nice study that extended our knowledge on employing new artificial intelligence-based inverse modeling methods to infer CO₂ fluxes from atmospheric observations. I have a feeling, the current content looks a bit “thin”, some advantages of the new method have not been clearly illustrated. Also, some detailed analyses should be added to make the presented results more sound. I would like to recommend it for publication after addressing the following issues.

Major comments:

1. The current results do not clearly illustrate the advantages of the GP/ML inverse method. Comparing the new method with old ones should help.
2. The presentation of results:

• Figure 2, I expect to see a time-series plot (if it can not be made for pixel scale, for regional scale also work), which can better show the performance of model optimization for prior/posterior and observations. Histogram plots for prior/posterior residuals (simulations minus observations) would also help.
• Figure S1 showing the spatial distribution of the posterior fluxes in the supplemental files can be moved to the main text.
• I expect to see the seasonal cycle (with a monthly time-step) of the prior/posterior CO₂ fluxes, for fossil fuels, fires, and NEE. These would help to see if the constrained fluxes can indicate the impact of COVID-19, wildfires, and seasonal anthropogenic emissions. Currently, we only see the results by season.

        3.  Is it possible to perform one-year more inversion? So then we can better understand the performance of the inversion model in revealing the impact of disturbance from COVID-19, wildfires, and droughts.

        4.  In Section 2.1, I see the boundary conditions were taken from GEOS-chem 4D-Var run at 4° ×5°, it is relatively coarse compared to the resolution of 0.5° ×0.625° for the inversion. I am not sure if this leads to some uncertainties for the inversion. Some higher-resolution BC, e.g., from CarbonTracker, might be better for the current inversion. Or some tests about the sensitivity of BC can be added.  

Minor comments:

1. Line 169, constraining-> constrain?
2. Section 2.3, I expect to see a spatial map showing the data coverage of OCO-2 and OCO-3 XCO₂ observations over the study area.
3. How to optimally determine the hyperparameters of the GP/ML model? It is not clear.