Gao et al. present a spatially gapless, daytime hourly tropospheric NO₂ vertical column dataset at 0.05° resolution over four representative polluted regions in Asia for 2019–2024. The product is designed to emulate what GEMS would have observed, free of the gaps caused by cloud contamination or retrieval failure. It is reconstructed by a deep-learning model: TROPOMI tropospheric NO₂ columns, ERA5 meteorological fields and EAC4 total column NO₂, together with land-cover and terrain data, serve as model inputs, while GEMS observations from 2023 are used as the training target. The dataset could be useful to researchers who wish to exploit GEMS NO₂ data but are constrained by its short record prior to launch, or by frequent cloud-induced data loss.

The subject is well suited to ESSD, and the manuscript is clearly structured and generally well written. I have a number of major and minor comments, listed below.

Major comments:

1. The abstract should be clearer. It is not clear from the abstract which variables were used as model input and which were used as the training target; please state this explicitly. Please also make clear that the dataset covers daytime hours only, and define what the EAC4 chemical prior is.
2. Section 3.1, line 248: how did you treat negative values in TROPOMI/GEMS/GOME-2? Negative retrievals are part of the random error distribution and cancel against positive noise during spatial or temporal averaging; discarding them gives an overall high bias. If negative retrievals were removed, please justify the filtering and quantify the resulting bias in both the training labels and the final product.
3. Regarding the gap filling: all your evaluation, including that for 2024, is performed over pixels where observations exist, so the gap-filled values themselves are never evaluated. This limitation should be stated explicitly in the text, as the reported metrics do not support claims about the quality of the dataset as a whole. In addition, EAC4 is the only NO2 input with complete coverage, so the gap-filled values presumably depend largely on it. Please quantify how strong this dependence is, discuss what happens if EAC4 is biased, and explain why a user should prefer your product over EAC4 itself in these regions.
4. Lines 483-485: you attribute the drop in model performance in 2024 to the model being unable to capture the inter-annual shifts in emissions. If so, how can we trust your predictions prior to 2023, when real emission changes were occurring?
5. Section 4.2: since GOME-2 provides long-term observations, please add an interannual evaluation against it. Does GOME-2 show trends and interannual changes similar to those in your product (at 09:30 LT from 2019 to 2024, Section 4.3)?
6. Section 4.3: you attribute all the interannual changes in the dataset solely to emission changes. Yet you also attribute the drop in model performance in 2024 to the model not capturing inter-annual shifts in emissions (Comment #4). These two statements are in conflict. Regarding these emission changes, do you have supporting evidence from other sources, such as government reports, bottom-up inventories, or surface NO2 observations, to support your statement. For example, you state that NP and KJ underwent abrupt and widespread emission reductions in 2024 compared to 2023 – is there independent evidence for this? Could meteorology (e.g., transport) also play a role?

1. The groups in Fig. 13 contain very different numbers of channels: 12 for static geography, 9 for transport, 8 for thermodynamics, but only one each for EAC4, SZA and TROPOMI. Since the zero-out perturbation removes a whole group at once, larger groups will produce larger deviations, so the finding that the three largest groups rank in the top three in every region may be an artifact of aggregation. This also explains the apparent conflict with Fig. 12, where EAC4 ranks in the top three for winter NP and for most months in KJ. Please consider normalizing by channel count or reporting per-channel importance, and revise Section 5.1 accordingly.

Minor comments:

1. Last paragraph of the Introduction: please clarify the spatial and temporal coverage of the dataset.
2. Line 158: what is the spatial resolution of the GOME-2 NO₂
3. Figure 11: which day is shown? It would be better if it were from 2024 instead of the training sample.
4. Figure 11: the prediction shows an anomalous feature along the border between South Korea and North Korea. Why?

The manuscript presents a gapless, 0.05° hourly tropospheric NO₂ VCD dataset covering four Asian pollution hotspots over 2019–2024, together with the deep learning framework used to construct it. The framework, PhysNorm-Net, couples a partial-convolution U-Net encoder operating on sparse daily TROPOMI NO₂ fields with a parallel ResNet “physics stream” that ingests 0.25° meteorological and chemical priors. This is a well-engineered study addressing the coverage gaps inherent in current satellite observations, and it is overall well-written. I hope the authors can address the comments below.

1. It is unclear how the predictions are validated over the regions where GEMS observations are absent (the “gaps”), which is precisely where the reconstruction adds value. I suggest a dedicated holdout experiment: artificially mask out blocks of valid GEMS pixels at inference, and examine whether the model’s filled values agree with the withheld observations, as a function of gap size. This would directly quantify the accuracy of the gap-filled portion of the dataset.
2. No GEMS observations exist before 2023, so GOME-2C is the only independent check on the 2019–2022. The sub-daily variability in the pre-GEMS years is never validated at any hour other than mid-morning (9:30). The manuscript should state this limitation explicitly in the abstract and conclusions, and ideally support the pre-GEMS diurnal cycles with independent hourly observations (e.g., Pandora/MAX-DOAS stations in the KJ and NP domains). Also, the agreement in Figure 5 is not visually convincing, in Figure 5a in particular, it is hard to conclude that the predicted spatial pattern is consistent with GOME-2C.
3. The 2023 data are randomly partitioned by date. Because both NO₂ and the meteorological priors are strongly autocorrelated at the synoptic scale, a random day-level partition places temporally adjacent, dynamically similar days on opposite sides of the split, so the test set is not independent of the training set in the sense the reported metric implies.
4. The systematic underestimation of the high tail concerns exactly the regime that matters most for the advertised applications, such as pollution episodes, health exposure, and policy assessment. Yet it remains unquantified.
5. “Hourly” is in fact daytime hourly. This should be clarified in the title, the abstract, and the dataset documentation.
6. EAC4 tcno2 is a total column, while the target is tropospheric. In addition, EAC4 is 3-hourly and linearly interpolated to 1-hourly, which smooths exactly the diurnal chemistry the model is meant to reconstruct. The paper should discuss the expected consequences of both choices.
7. The zero-out perturbation used for feature importance conflates shared variance among correlated inputs (e.g., q and T at four pressure levels each), so the reported rankings should be interpreted cautiously.
8. Cross-sensor consistency is not discussed: the labels are GEMS, the high-resolution observational input is TROPOMI (~13:30 LT), and the historical validation is GOME-2C (~09:30 LT). These retrievals differ in AMF formulation, a priori profiles, and resolution, and those systematic differences are silently absorbed into the learned mapping. A brief inter-sensor bias assessment over coincident scenes would substantially strengthen the validation chapter.