This manuscript introduces a new neural network (NN) based approach for correcting estimates of Arctic downwelling shortwave radiation (DSR) and, subsequently, Arctic net surface radiation (NR) from CERES products, cloud properties, and ancillary atmospheric and surface properties.  The algorithm improves DSR estimates relative to surface flux observations, especially in cases where CERES underestimates cloud fraction relative to a recently developed cloud product that combines active and passive observations.  NR is also improved, though to a lesser degree, likely due to the approach neglecting the varying influences of downwelling longwave radiation (DLR) on NR.  There is value to producing more robust DSR estimates that incorporate improved estimates of cloud fraction from active sensors as well as dependences on other atmospheric and surface conditions.

My primary concern with the study is that the NN approach introduces a disconnect between the final DSR and NR estimates and the physics that modulated them.  While the multi-variate NN captures nonlinear relationships and includes additional factors that modulate surface radiative fluxes, it masks the precise physical relationships that led to the results.  One clear example of this is the fact that NR is estimated from DSR without accounting for cloud or atmospheric influences on longwave radiation.  Presumably the NN captures some of the longwave effects through covariances between DSR, DLR, and other regression variables but it cannot make up for the lack of information provided by longwave radiative transfer calculations.  The direct physical connection between inputs and simulated fluxes has considerable value for many atmospheric process and climate applications, so it is not clear how this product could be used in those contexts.

Considering both the value of the analysis and the associated concerns, the paper may be suitable for publication after major revisions to better explain which applications the NR product may be address and responding to the following comments.

Specific Comments:

1. The introduction is poorly structured and difficult to follow at times. For example, the sentence about limitations in detecting low-level Arctic clouds on Line 55 doesn’t seem to directly follow from the sentences preceding it.  The paragraph that begins on Line 64 discusses the PRP method needs to be better motivated and connected to the goal of estimating surface radiation.
2. Line 91: For the general reader, it would be useful to explain what an Angular Distribution Model is and how it works.
3. Equation 7 and the sentences that follow it should appear immediately following its reference in the text on Line 199.
4. Line 221: delete ‘budget’
5. Section 4.4: it is important to note here that deriving NR strictly from DSR, albedo, day length, vegetation cover, temperature, and humidity severely limits the utility of the estimates for process or climate studies. This method essentially assumes that longwave cloud radiative effects can be parameterized from these variables and allows the NN to invent that relationship.  Since these relationships are far from unique, physically interpreting the causes of variation in the resulting NR is largely impossible.  Even if less accurate based on surface measurements, the independent estimates of DSR and DLR in the CERES product contain more information to enable climate forcing and feedback studies.  To provide a concrete example, studies have shown that trends in wintertime surface radiation play a role in ice loss the following year.  How can an algorithm that derives NR from DSR provide any insight into the dark months in the Arctic?  The utility of the NR estimates from the ANN approach should be clearly stated as well as the limitations relative to the original CERES product that covers all seasons.
6. Figure 10: following from the previous comment, comparisons of NR are only presented for the daylight months of April – September. Obviously NR cannot be estimated from DSR in winter months but switching between the ANN estimates here back to CERES products in winter would introduce a discontinuity in the physics for analyses spanning the whole year.  Again, the utility of the NR estimates introduced here should be clearly stated.
7. Section 5.2 seems out of place in this study. It is not clear how this discussion of radiative kernels relates to the methods introduced in the paper other than to show that radiation dependences on clouds are nonlinear and complicated by atmospheric and surface conditions.  No additional insights into cloud effects or feedbacks are gained from this discussion so I would suggest removing or significantly abbreviating it.
8. Line 471: ‘accuracy’ should be ‘accurate’
9. Line 487: The extended model doesn’t avoid LW uncertainties, it introduces them but not explicitly accounting for cloud impacts on DLR. This sentence should be modified.
10. Line 500 - 501: In my opinion, applications to climate model evaluation and monitoring Arctic amplification are severely impeded by the lack of explicit consideration of DLR in the NR estimates provided here. It is not clear how one would interpret model biases relative to the output generated from this algorithm or whether the method would adequately capture trends in NR associated with Arctic amplification.  It is also not clear how well these methods would extrapolate to other regions where variability may be more closely tied to longwave radiation than shortwave.

1.    A variety of cloud properties influence cloud radiative effects, including cloud optical thickness, particle effective radius, and several others. In this manuscript, however, cloud fraction is treated as the key variable. The authors are encouraged to clarify the reasoning for prioritizing cloud fraction over other potentially important cloud parameters.

2.    Using the improved DSR to estimate NR is a good idea, and the strategy of bypassing the longwave component by introducing other parameters is interesting. The rationale for selecting these specific variables is not fully clear. The authors are encouraged to explain why these variables were chosen.

3.    In estimating NR, the authors incorporate multiple parameters, some derived from CERES and others from various independent products. Because these datasets differ in spatial resolution, it is not clear how the authors handled the mismatch in resolution when constructing the training dataset and matching predictors with the target variable. The authors should clarify the preprocessing steps used to ensure spatial consistency across datasets.

4.    In Figure 9, the caption refers to “Validation of DSR,” but this figure is validating NR. Please revise the caption accordingly.

5.    In the Ground-measured data description section, the authors should provide appropriate references for all data sources used for the observational stations.