In this manuscript, the authors apply deep learning to reconstruct historical global surface temperature fields, with a specific focus on Antarctic climate dynamics. While this topic is relevant to research on climate change impacts and attribution, I have several concerns regarding the methodology, the robustness of the validation, and the demonstrated utility and uniqueness of the final products. Additionally, various statements throughout the text require further clarification. Please find my detailed comments below.

1. To my knowledge, the reconstruction framework based on a partial convolutional neural network (PConv) was first applied to historical climate field reconstruction by Kadow et al. (2020). The authors follow a nearly identical methodological approach. The authors must explicitly acknowledge Kadow et al. (2020) in the methodology section. This will help readers accurately contextualize the foundational literature and better understand the specific novel contributions of this work.

2. Following the previous point, the authors should at least use the latest dataset built by Kadow et al. (2020) as a benchmark to evaluate the reliability and added value of the newly proposed dataset. Also, regarding the uniqueness of the proposed dataset, it would be interesting and necessary to see how the latest Kadow et al. (2020) product performs in the Antarctic region compared to this study.

3. The evaluation and validation sections need to be strengthened. Currently, it is difficult to distinguish the unique strengths and irreplaceability of these new datasets from those of existing reference datasets solely on the basis of global/regional temperature trends. Providing a clearer demonstration of where and why this dataset outperforms existing ones would greatly improve the manuscript.

4. I recommend moderating some of the claims regarding the capabilities of AI methods. While powerful, some statements about AI's strengths and the overall contribution of the study come across as overconfident and could be nuanced to reflect limitations of methods and datasets.

5. The study provides two reconstruction products, but it is not entirely clear which one is considered superior or more reliable. A more solid comparison between the strengths and characteristics of the two products would be helpful. Furthermore, providing explicit guidance on which dataset users should select for specific use cases would add practical value to the paper.

Specific Comments:

L47-50: The authors mention that the "propagation of observational errors", "inconsistencies among reanalysis products", and "harsh environmental conditions" affect existing statistical reconstruction methods (PCA, EOT, DINEOF). While true, it is worth noting that AI-based methods often struggle with these exact same regional challenges. I suggest softening the critique of traditional methods here.

L58: Please clarify what is specifically meant by "extreme geographical conditions" in this context.

L98-99: Please remove the claim regarding interpretability. PConv models are generally not recognized for their physical interpretability.

L104-109: The introduction of the C-MST3.0 dataset is somewhat confusing, given that the reconstruction primarily relies on C-LSAT2.1. Consider revising this section to streamline the data origins.

L136: Please elaborate on what the "influence by features at the edges of missing data" specifically entails in this context.

Figure 1: This schematic appears to share strong similarities with the framework figure in Kadow et al. (2020). Please ensure appropriate citations or copyright permissions are included if it was adapted. Additionally, the numbers at the bottom of most boxes are somewhat confusing and could benefit from an explanation in the figure caption.

L149-150: Please specify exactly what is meant by "the AI reconstruction follows a similar approach." Similar to what?

L168: Provide more details on the regridding process. What exact resampling method was used?

L205-211: The authors claim that the AI reconstruction performance under the Merge-E mask is better than that of Merge-H, attributing the smoother reconstructed fields in Antarctica to the AI being "influenced to some extent by the colour, texture, and style features at the edges of missing regions." This conclusion is debatable. Because ERSST is a spatially complete, interpolated dataset, it lacks the vast oceanic gaps present in datasets like HadSST. Therefore, the "smoother" performance observed in Merge-E is likely not due to PConv's enhanced inpainting capabilities, but rather to the statistical interpolation already performed within the ERSST dataset itself. Please revisit and clarify this discussion.

L215-217: Highlighting the single-year result of the 1877 El Niño is a good visual check, but it is not sufficient to claim "robust performance and spatial consistency." Moreover, to my knowledge, the infilled HadCRUT5 dataset can also represent this 1877 event. Additional spatial validation is needed to support this claim.

L224-253: The detailed analysis comparing Merge-H and Merge-E feels somewhat redundant. Since ERSST is already reconstructed and spatially complete, applying AI reconstruction to it does not add much novel insight. The authors might significantly condense this section to simply justify the choice of Merge-H rather than present it as a core comparative result.

L253-254: It is difficult to visually distinguish among the various products. Consider analyzing and plotting the residuals (differences) instead, which would make the variations much clearer to the reader.

L255-256: Only one out of the four AI reconstructions shows a GWL > 1.5. Given this, the conclusion drawn here feels too confident. It would be appropriate to discuss the spread/uncertainty among the models.

L259-270: Inferring historical climate dynamics (i.e., ENSO) purely from the Global Mean Temperature (GMT) time series in Fig. 3 seems to be an over-interpretation. The claim that this "further demonstrating the robustness and reliability of the AI framework" feels overstated.

Tables 2 & 3: Please add the other reference datasets to these tables for comparison. Without them, it is difficult to objectively assess the performance and reliability of the AI reconstructions.

L285-286: It is difficult to visually confirm in Figure 4 that "the differences between different reconstruction results decrease  markedly." Consider specifying timeframes and supporting them with a quantifiable metric.

Figure 5: Please improve the visual clarity of this figure. The lines overlap significantly, making it difficult to distinguish the differences between the datasets.

L361-362: There appears to be a contradiction here: the text states that CMIP6-AI performs better than 20CR-AI, but Figure S6 seems to show the reverse. Please verify and clarify.

L368-371: Please rephrase this statement for clarity. It is not entirely clear how the reference datasets are being quantitatively or qualitatively compared with the station data.

L389-390: Please specify the exact statistical method used to measure the significance here.

L423-424: The conclusion that the products show high spatiotemporal consistency is not fully supported because it has not been benchmarked against other references or methods. Basing this conclusion solely on GMT and regional trends is clearly insufficient.

Section 5 (Limitations and future perspectives): As the ESSD journal is data-focused, this section should focus specifically on data limitations, uncertainties, and the irreplaceability of the generated datasets. Broad discussions of the future of AI methodologies are somewhat outside the journal's primary scope and should be minimized.

Overall, while the core concept of this study is compelling and timely, the manuscript's current presentation, the robustness of its validation methods, and the understanding of the underlying datasets do not yet meet the necessary threshold for publication.

This manuscript presents a deep learning approach based on partial convolution, trained using 20CR reanalysis data and CMIP6 historical simulations, to reconstruct global monthly gridded surface air temperature from 1850 to 2024, with a particular improvement over the Antarctic region. Overall, the manuscript is well-structured, the methodology is clearly described, and the design of the reconstruction framework as well as the selection of sea surface temperature datasets are appropriate. The reconstructed results are validated against multiple observational datasets and Antarctic station records, demonstrating good consistency. The resulting dataset is valuable for studies of global temperature variability, particularly for Antarctic climate analysis and global temperature change assessments, and is generally well aligned with the scope of ESSD as a data journal.

It is also noteworthy that Dr. David Bromwich, an expert in Antarctic temperature research, pointed out during the public discussion that the Antarctic temperature reconstruction presented in this study shows certain advantages compared to state-of-the-art reanalysis datasets (e.g., ERA5).

I have no objection to the publication of this manuscript after minor revisions. The following comments mainly concern aspects such as figure presentation and result interpretation, which should be further improved by the authors during revision:

1. In Figs. 2, 4, and 5, the font sizes of the axis labels and tick marks are relatively small and difficult to read. It is recommended to increase the font size to improve readability. In addition, the legend in Fig. 6 appears somewhat blurred; improving the resolution of this figure would help enhance clarity.

1. Figure 2 presents reconstruction results for four representative months, but the manuscript does not explain the criteria used to select these months. It would be helpful if the authors could briefly clarify whether these months represent different observational coverage periods, climate conditions, or other considerations.

1. Figure 3 shows the global annual mean temperature time series. Since the reconstructed series is visually very similar to the CMST3.0-Imax dataset, the contrast between the curves could be improved. For example, the CMST3.0-Imax series could be plotted with a thicker solid line to make the comparison clearer.

1. In Table 1, the warming trends for Merge-H 20CR-AI and Merge-H CMIP6-AI during 1850–2024 are very similar (both 0.064 ± 0.006 °C per decade), yet the corresponding global warming levels (GWL) differ substantially (1.45 and52 °C). However, lines 255–256 state that “The GWL reconstructed by CMIP6-AI reaches 1.52 °C, which is slightly higher than the 1.5 °C estimated from 20CR-AI.” The authors may wish to check whether the GWL value reported in Table 1 is correct or if there is a discrepancy between the table and the text.

1. The manuscript notes that observational records in Antarctica are sparse prior to 1961, making direct validation difficult. It would be helpful if the authors could briefly discuss the limitations or uncertainties associated with the reconstructed temperature fields during this early period, so that potential users of the dataset are aware of these limitations.

1. In line 412, the manuscript refers to “optimal interpolation, EOF, or spatially weighted averaging”. The term “EOF” in this context may need clarification. If the authors intend to refer to the empirical orthogonal teleconnection method, it may be more appropriate to use “empirical orthogonal teleconnection (EOT)”.

1. There are a few minor language issues throughout the manuscript. For example, in line 10, the comma in “substantial missing information in global ST datasets, remains a major source of uncertainty” should be removed. In line 63, “observation-based regional temperature fields often hard to capture” could be revised to “often fail to capture”. A careful language editing pass would further improve the readability of the manuscript.

Please find attached a document containing my detailed comments on the manuscript.