Major comments:

Section 2.2.2: Six 2D and 3D indicators are selected as predictor variables for CUHI, but there can be more indicators. Could authors justify why these indicators are used? A review on morphology variables used in previous regression/ML methods is needed.

Section 2.3.1: How many HW days are found based on the criteria used in this study? This information can be put in Figure 2 to better illustrate the length of HWs.

Section 2.3.3: The training process of XGBoost model requires more details. What data is used as training, validation, and test set, respectively? How is the model performance evaluated? This is the major flaw because the results in Fig. 5 and Fig. 6 will be significantly affected by the model performance.

Section 2.3.4: How did authors select study areas for ENVI-met simulations? And what are the values used for various thermal properties in the model setup?

Line 177: the larger nighttime CUHI than daytime CUHI shall be better explained. there have been many studies in the literature, and it will be good to have at least some comparisons against CUHI during HW at different cities.

Line 189-196: The explanation here relies on visual interpretation of Figs. 3 and 4. I think this part can be removed as Fig.5 shows more reliable statistical analyses.

Fig.5: Are these results from XGBoost model? How is the model evaluated? For daytime results, the correlation value is small for all indicators except for BCR, which is only about 0.3; This seems to suggest that model performance is bad, or no single indicator is powerful enough to explain the CUHI. For nighttime results, many 3D indicators have coefficients very closed to SVF, and thus it is hard to argue that SVF is the dominant factor. The results can be changed with slight modifications of the data or training processes. Without rigorous model validation, the SHAP results in Fig.6 are less meaningful. 

Fig.8: the derivation and meaning of PDP plots shall be elaborated for general readers not familiar with this method. Current discussions related to Figure 8 are hard to understand.

Fig.9: How did the authors modify the physical domain to have different SVFs, increase building height or reduce road width? Is such increment or decrement uniform across the entire domain?

Figs. 10 and 11: After changing SVF in the ENVI-met domain, authors only analyze the temperature at the central point of the domain, this is too simple. In fact, using ENVI-MET at 1 neighborhood with different SVFs to demonstrate that temperature will change differently from NHW to HW does not sound convincing or necessary. 

Section 4: the discussion section focuses on analyzing the impact of wind on CUHI. However, the correlation is very weak. In addition, this part seems to deviate from previous correlation and SHAP results. From my perspective, authors seem to combine too many methods (XGBoost, ENVI-met, and correlation with wind) to explain CUHI change under HW, and this paper lacks a good organization and logic flow. After reading the paper, I am not sure what authors aim to address, and what are the key findings.

Minor comments:

Fig.1 caption: "Overview of study area" is repeated. 

Line 167: remove "the next day" as this is a averaged diurnal cycle

Fig.3 caption: only (a) and (b) sub-figures; and I suggest authors to add the different in CUHI between HW and NHW to better illustrate the distribution of CUHI change

Major comments:

Section 2.3.1: The HW definition requires stronger justification. Specifically, why did the authors decide to use reference stations to define HW and why was the threshold set to “more than two reference stations”?

Section 2.3.3: More details on the training/validation processes of XGBoost are needed. How were the collinearity among morphological indicators (e.g., FAR and BCR) treated in the ML models? More detailed explanations of SHAP and PDP methods would improve reader comprehension of the results in Figure 6-8.

Section 3.2: Before presenting the analysis of Figures 6-8, there should be at least one figure showing the model performance of XGBoost, as the validity of these results strongly depend on the model’s predictability of CUHII.

Line 255-257: This summary largely repeats content in line 236-239. The authors should streamline the conclusion from figure 8, e.g., focus more on the nonlinear modulation.

Section 3.3: The scenario setup requires clarification. How were the uniform SVF values applied across the entire domain in scenario II and III? Does scenario I have spatially heterogeneous SVF values? If so, the rationale for using uniform values in scenario II and III needs justification. Currently it is difficult to interpret spatial changes in Figures 10-13 with most discussions focused on the central point.

Section 4: While the wind-CUHII relationship is worth discussing, the analysis should emphasize how urban morphology modulates wind patterns to be tightly connected with the main theme of this work. The current presentation of Figures 14-16 lacks clear connection to morphological controls, making it difficult to identify the key messages.

Minor comments:

Figure 3: There are (a)-(f) in the caption but only four subplots are presented.

Figure 6: It would be easier to interpret if the authors can group 2D and 3D indicators (e.g.presenting all 2D indicators in the first 6 rows, followed by 3D indicators).

Figure 9d: Does the ‘simulation accuracy’ refer to scenario I?