This paper presents a deep-learning based approach for 2D surface flood modelling that builds on a previously existing model, SWE-GNN. It proposes an alternative model that overcomes the need of a numerical solver to determine initial conditions and includes other novelties to improve the speed and generalization of the model. The paper demonstrates that the model can benefit from fine-tuning to generalize to new case studies. In my opinion, the research is interesting and is very well presented. I am therefore recommending the paper for publication, after minor revisions.

General comments

The paper proposes a model for flood modelling and is applied only to a specific case of floods; dike-breach floods. It would be nice to have a section in the discussion on how the model could apply to other kind of floods, such as fluvial and pluvial floods for example.

Specific comment

L84 – L85: For reproducibility, I would suggest clarifying how a mesh is classified as being too small and how the resolution of the fine mesh is chosen.

L93: It might be better to define what epsilon is here rather than in L104.

Eq. 3 and 4: I think that it might be useful to have the clarification of what h_di and h_si are.

L190: I would suggest replacing ‘training simulations’ by ‘training data’ here for clarity.

L193: You should explain that water level is the sum of the water depth and elevation of the cell as this is might not be straightforward.

Section 2.5: Clarify that O is the output dimension.

Table 1: To which resolution of the mesh does the edge length refer?

Section 3.1: I believe a concise explanation of how the Manning’s n coefficients are defined for the synthetic dataset would be helpful. I’m curious to know if these coefficients are spatially variable as the model might face challenges in extracting meaningful insights from them if they were spatially uniform.

L270-275: You provide a nice and clear explanation of CSI. However, you consider 2 thresholds (0.05 m and 0.3 m) while you explain in L273-275 that the flooded and non-flooded cells are distinguished from another in TP, FP and FN. However, this isn’t entirely accurate as cells with water depths below 0.3 m could still be considered as flooded. I would try to be more accurate and write for example: ‘TP are true positives, i.e. number of cells where both model and simulations predict water levels above the threshold value’. The same applies for the description of FP and FN.

L285: I would suggest adding a sentence stating that you are assuming the enhanced SWE-GNN also outperforms the other models, given that the original SWE-GNN outperforms them and despite the added modifications.

L305: Do you know why the MAE of h increases? Could it be due to error propagation throughout the simulation? If you have any insights into the cause of this increase, it might be worth adding it.

Fig. 9 and Fig. 10: Consider using different color scales as it is difficult to visualize the differences in water levels as it is. Also, could you clarify which mesh (e.g. finest mesh) is shown in the figures?

Fig. 10: It is really nice that you detail what the negative and positive values correspond to in the captions of Fig. 9 and Fig. 10. However, I would recommend revising the caption in Fig. 10 to enhance clarity, e.g. ‘positive values indicate that the model estimates later arrival times than the numerical simulation, while negative values indicate that the model predicts earlier arrival times’

Technical corrections

Throughout the manuscript (e.g., L14, L51, L223, caption of Fig.5): Uncapitalize ‘the’ in ‘The Netherlands’

L178: ‘via edges directed towards’ rather than ‘via directed edges towards’

L292: Clarify ‘is comparatively faster than the numerical model’

Throughout the manuscript: Add a spacing between the numbers and the ‘m’ for meters

L302-303 and 305: Replace the four ‘in correspondence of’ with ‘occur at’, ‘occur simultaneously to’, ‘near’ and ‘at’ respectively

Fig. 8: You might want to keep the notation of CSI_0.05 and CSI_0.3 like in the manuscript, i.e. add the ‘m’

Table 2 and 3: Write the units as [10-2 m] and [10-2 m2/s]

L330: Typo in ‘dataset’

L365: Typo in ‘dependent’

L393: ‘analyze’ rather than ‘analyse’ to keep consistency with the previous sentence

L466: Remove ‘the’ in ‘comes from the an increase’

Review of "Multi-Scale Hydraulic Graph Neural Networks for Flood Modelling"

This paper presents a multi-scale, hydraulic-based graph neural network model (mSWE-GNN) designed to enhance flood modelling. Building on previous models like SWE-GNN, the authors introduce novel elements such as rotation-invariant inputs, ghost cells for boundary conditions, and a multi-scale architecture to simulate flood events over variable topographies and unseen meshes. The paper includes extensive theoretical explanations and benchmarking evidcence that demonstrate the model's effectiveness across different flood scenarios. Overall, this research is timely and relevant, contributing meaningfully to the field of flood modeling with machine learning. I support the publication of this paper following some revisions.

General Comments

The model presented in this paper, mSWE-GNN, demonstrates impressive capabilities in simulating dike-breach flood scenarios specifically. However, its applicability appears limited to this particular type of flood. Given the diversity of flood types—such as fluvial and pluvial floods—extending the model's capabilities or discussing potential adaptations for these other scenarios would significantly enhance its relevance and utility. A broader application across flood types would open up further opportunities for practical deployment and highlight the robustness of the proposed multi-scale architecture in different hydraulic contexts. Including a discussion on how the model might be adapted for other flood types would be a valuable addition.

However, the paper limits its application to dike-breach floods and does not explore other flood scenarios. Adding a section discussing potential adaptations for other flood types, such as fluvial or pluvial floods, would strengthen the paper’s relevance. Additionally, the inclusion of a mass conservation term in the loss function, though discussed in the appendix, warrants further investigation, particularly regarding its influence on model stability.

Specific Comments

• L89-92: For clarity and reproducibility, it would be beneficial for the authors to provide additional detail on the criteria for mesh size. Specifically, a brief explanation of how mesh resolution is chosen and classified as "too small" would enhance reproducibility.
• Eq. 3 & Eq. 4: It may be helpful to specify definitions for hdih_{di}hdi​ and hsih_{si}hsi​, as these variables are key in understanding the propagation rule and its application.
• Boundary Condition Section (2.3): This section introduces ghost cells effectively but lacks explicit examples or visual representation of how they are implemented for various boundary types (e.g., inflows and outflows). Including this could aid in comprehending the technique's generalizability to different boundary configurations.
• Fine-Tuning Evaluation (Section 4.2): While the fine-tuning approach is well-motivated, it would be helpful to elaborate on the potential for overfitting given that a single simulation is used for fine-tuning and testing. The authors might explore this limitation and offer suggestions for future validation.
• Discussion on Physical Constraints (Appendix B): The addition of mass conservation loss is an interesting consideration. However, it would be helpful if the authors could provide further analysis or reasoning on why this term, although theoretically sound, did not yield a consistent improvement in testing. This section could benefit from a comparison of other physical constraint approaches in neural networks.

Figures and Tables

• Figures 9 and 10: The flood arrival time and water depth predictions presented are informative, but some color scales may not adequately differentiate between close water depth levels. Adjusting these color scales would make the figures more accessible and highlight differences more clearly.
• Table 2: To support interpretability, clarify which mesh resolution is referred to in this table. Also, in Table 3, presenting units consistently as [10−2 m][10^{-2} \text{ m}][10−2 m] and [10−2 m2/s][10^{-2} \text{ m}^2/\text{s}][10−2 m2/s] improves clarity and aligns with standard formatting practices.
• Appendix Figure A1: This figure is valuable in understanding unit discharge modeling, but it may benefit from a clearer representation of errors by using varying scales or alternative visual indicators for over- and under-predictions.

Technical Corrections

• L302 & L305: The phrase “in correspondence of” could be replaced with more direct alternatives such as "occur at" or "align with" for clarity.
• Throughout manuscript: Maintain consistent formatting of units, particularly with spacing between numbers and "m" for meters.
• L466: “comes from the an increase” should be corrected to remove the redundant article “the.”

Conclusion

In summary, the paper presents an interesting deep learning model with clear advancements over existing approaches in flood modeling. The mSWE-GNN’s ability to generalize and efficiently process multi-scale data is convincingly demonstrated. Nevertheless, a limitation of the current model is its focus on simulating dike-breach floods alone. Given the range of flood types—such as fluvial and pluvial floods—future work extending this model’s application to other types would significantly increase its value and utility. Including a discussion on potential adaptations for different flood scenarios would further enrich the paper, offering insights into the generalizability and versatility of the multi-scale approach.