General comments

This paper by Duke and co-authors describes a neural network approach to interpolate observations into a gridded pCO₂ data product over the period 1998-2019 at 1/12° resolution specifically in the Northeast Pacific. After presenting the specific circulation and hydrographic features of their area, the authors detail the methodological choices they made in developing the NN and advise the reader on specific training steps to improve performance. They then go on to discuss the wider scientific applications of such a product.

Overall, the paper is very well written, nicely structured and easy to read. The figures could be larger to improve readability, but figures and legends are still clear. The level of detail in the methodological and NN development sections is much appreciated, as it is rare to find such detail in oceanographic research papers. In particular, the decision not to normalise all your inputs and to use dynamic provinces is very interesting. Also all the developments you have done on the internal division data ratio, which could be applied to many other uses.

A few specific comments remain, as detailed below:

Specific comments

• It would be nice to further highlight the potential uses of your data product and/or your method outside the traditional straightforward pCO₂/FCO₂ observing community (i.e. modellers, determination of climate indices, ...).
• Have you considered comparing your FCO₂ with other products: e.g. SeaFlux (Fay et al., 2021)?
• End of section 2.1 and Table 1: there seems to be some confusion here about what you have used for what. You mention the xCO₂ data produced by NOAA, but in Table 1 you cite Lanschutzer? As in the next sentence you mention the pCO₂ climatology by the same authors, restructuring the sentences if this is not a confusion would be very useful.
• Table 1: for SST you kept the 1/20° resolution and did not aggregate to 1/12°, unlike the others?
• Figure 2c: Is your validation data set representative enough? It does not cover your whole range?
• End of section 2.5: You average the outputs of the 10 NN. You do this directly, but have you tried to give the median +/std as this is also useful information.
• Line 210: "12.9 +:-1.1 µatm" : I'm not sure where this number comes from.
• Line 332: flux densities as high as 3.6, but your figures end at a maximum of 3?
• Lines 394-395: provide some insight into how.
• In this section you compare an atmospheric xCO₂ with an increase in oceanic pCO₂. Can't you convert the atmospheric xCO₂ to pCO₂ to compare things in the same ranges/units?
  
  pCO2ATM = [PT - (RH/100) × PH2O] × xCO2ATM
  
  where PH2O is the water vapour pressure at atmospheric temperature for xCO2ATM (in atm) calculated according to Dickson et al. (2007), RH is the relative humidity (in %) and PT is the total atmospheric pressure (in atm). If some of these are missing, you can get them from products (SeaFlux mentioned above). Otherwise you are comparing things that are not directly comparable.
• Line 437: detail how you got the vertical velocities with Ekman pumping

Technical corrections

• Abstract: For readers not familiar with your particular area, it would be better to introduce the fact that your area is a sink earlier.
• Line 58: "The estimated long-term trend in surface ocean pCO₂ appears to be increasing at less than the atmospheric rate." Less what? rate, amount of increase?  Rewrite to clarify: "at a slower rate" or "less than the atmospheric rate of increase".
• Line 119: fCO₂ has been converted to pCO₂: how, give equation.
• Figure 3: Add some metrics to the graphs (from the text)
• Line 150: "changing the internal division ratio", what did you end up using (see where you give more details).
• Line 424: "could be partially tied to if" missing word/rewording
• Figure 7: the black line is not really black.
• Supplementary Figure 1 a&b: useful to superimpose bar charts of the whole dataset behind to further demonstrate how your subsampled data is representative.

Summary:

Duke and coauthors use a two-step cluster–regression method to map surface partial pressure of carbon dioxide (pCO₂) in the Northeast Pacific Ocean. Their approach is novel in that they grid pCO₂ observations and produce pCO₂ maps with high spatial resolution (1/12°); in doing so, they offer insightful observations about optimal model parameters and regional driving factors of CO₂ flux variability. This work represents not only a useful product for investigating surface carbonate chemistry in the Northeast Pacific (NEP), but a valuable roadmap for increasing the spatial resolution of observation-based surface ocean pCO₂ products.

This manuscript is very well-written and clear to follow. I was especially impressed with the analysis surrounding the training of artificial neural networks with progressively finer resolution, and the critical nature of the training/evaluation data split in these instances. The examination of driving factors of CO₂ uptake variability and the effects of marine heatwaves is interesting and will be beneficial for researchers seeking a region-wide carbonate chemistry context for the NEP. I detail a few general and line-specific suggestions below, but overall support the acceptance of this manuscript.

General suggestions:

The conclusion that the training data to internal evaluation data ratio should be optimized and likely increased toward finer resolution grids will be extremely valuable as global-scale observation-based pCO₂ data products with finer than 1° resolution are beginning to be produced. In that context, it may be helpful to expand upon the statement at the end of section 3.4 that this result “creates a precedent for stepping to a higher resolution product with nearly no loss in performance”. How might you envision that higher resolution step being taken at a global scale? What are some important considerations and potential pitfalls when taking this approach beyond the NEP? Any thoughts about increases to the temporal resolution?

One limitation of the validation performed here is that the statistical metrics represent the ability of the ANN-NEP procedure to estimate pCO₂ only at the spatiotemporal grid cells where observations are available. This may mask location-specific seasonal biases, especially at high latitudes where wintertime observations are likely not as plentiful. In lieu of a comprehensive model simulation experiment to evaluate these unquantified biases, this consideration may warrant some discussion in section 3.2 or elsewhere.

A figure displaying the most frequent occurrence of each SOM province over the timeseries would be informative. As an additional suggestion for future work: to reduce the discontinuities at the borders of biogeochemical provinces it would be interesting to explore soft clustering approaches in addition to hard clustering like SOMs. Soft clustering approaches provide probabilities for each clustered grid cell, which can be used as weights to blend pCO₂ predictions across clusters.

Line-by-line comments:

Line 85: It would be valuable to articulate why the coastal ocean was excluded in this study.

Lines 142–143: It isn’t immediately clear why choosing not to normalize predictor data implicitly weights the SOM predictors toward the pCO₂ climatology. Is it related to the relative range of each chosen predictor?

Lines 151–152: I don’t understand what is meant by “we introduced each predictor variable again after deseasonalizing”. Can this be explained more clearly?

Lines 280–281: Very interesting and insightful conclusion!

Lines 455–456: It would be good just to clarify in this sentence that “stepping to a significantly higher spatial resolution” refers to a higher resolution “than typical observation-based pCO₂ products (1/4° or 1° resolution)” or something along those lines.