A Physics-Informed, Machine Learning Emulator of a 2D Surface Water Model: What Temporal Networks and Simulation-Based Inference Can Help Us Learn about Hydrologic Processes

Round 1

Reviewer 1 Report

After minor corrections can be accepted.

Comments for author File: Comments.pdf

Author Response

R1 Comments copied from those in marked up PDF

We thank the reviewer for the helpful comments.  We have extracted the comments from the PDF and respond below in italics.

Title is too long pls reduce it to 2 line Max.

We appreciate the suggestion. However, we think the title is descriptive and as it’s a two-part title, not too long.  Additionally, there is no title length requirement in Water’s Instructions for Authors: https://www.mdpi.com/journal/water/instructions

Line 57: Literature needs to extended more...

We thank the reviewer for this comment.  We added additional references in the revised manuscript; however, as our approach is quite new there are few references on ML emulators of distributed hydrologic models.

Lin 195: Do not need to present all figures pls select some important figures and use tables if you need more details

We streamlined the discussion and removed three figures from the manuscript.

Line 397: Pls present the relationship between your results and literatures of your article.

We thank the reviewer for this comment.  We have tried where possible to include connections to prior work.

Reviewer 2 Report

The authors of the paper “A physics-informed, machine learning emulator of a 2D surface water model: what temporal networks and simulation-based inference can help us learn about hydrologic processes” investigate the solution of the Tilted V catchment benchmark problem with the use of 6 ML models that are fed by a specific hydrologic model (for different ensemble members and sets).

The conducted work is very well structured and organized and tries to deeply foster ML in hydrology and the coupling of ML with spatial distributed models. Although the research focuses on a very specific audience, it’s very promising research and I fully enjoyed reviewing it.

Some minor revisions have to do with the following:

The accuracy of the paper outputs is based on the SBI approach, thus the accuracy of the method (in hydrology) should be further highlighted.

Figure 4 has the same colours with Figure 3, nevertheless they present different things. I propose that RMSE (Figure 4) should be somehow presented differently (maybe as a Table?) so the outputs to be better depicted. The same with figure 12.

In Figure 5 (legend) and in line 227, the authors speak about a shaded area between the plots, however this area is not visible. Thus, please redraw Figure 5.

Figure 5. The authors should indicate what the upper and lower dotted lines symbolize.

Lines 230-231: The authors state “This result is somewhat surprising given that the CNN2D_A1 model excludes pooling.” In which way the specific model excludes pooling (apart from just mentioned it) and how this is demonstrated in Table 5 (Annex)?

Author Response

The authors of the paper “A physics-informed, machine learning emulator of a 2D surface water model: what temporal networks and simulation-based inference can help us learn about hydrologic processes” investigate the solution of the Tilted V catchment benchmark problem with the use of 6 ML models that are fed by a specific hydrologic model (for different ensemble members and sets).

The conducted work is very well structured and organized and tries to deeply foster ML in hydrology and the coupling of ML with spatial distributed models. Although the research focuses on a very specific audience, it’s very promising research and I fully enjoyed reviewing it.

We thank the reviewer for recognizing the contributions of this work, and we are glad they enjoyed reading and reviewing it.  We reply to each comment below in italics.

Some minor revisions have to do with the following:

The accuracy of the paper outputs is based on the SBI approach, thus the accuracy of the method (in hydrology) should be further highlighted.

Thanks for the comment, we have significantly revised sections 2.5 and 3.4 to be provide more details and better highlighted the SBI approach in the revised manuscript.

Figure 4 has the same colours with Figure 3, nevertheless they present different things. I propose that RMSE (Figure 4) should be somehow presented differently (maybe as a Table?) so the outputs to be better depicted. The same with figure 12.

Thanks for the comment, we changed the color scales for these figures.

In Figure 5 (legend) and in line 227, the authors speak about a shaded area between the plots, however this area is not visible. Thus, please redraw Figure 5.

We thank the reviewer for catching this. The loss of shading was a result of typesetting for the review manuscript, we will make sure this figure is clearly represented in the final version.

Figure 5. The authors should indicate what the upper and lower dotted lines symbolize.

We thank the reviewer for this comment, we have indicated this in the caption of this figure.

Lines 230-231: The authors state “This result is somewhat surprising given that the CNN2D_A1 model excludes pooling.” In which way the specific model excludes pooling (apart from just mentioned it) and how this is demonstrated in Table 5 (Annex)?

We have added additional detail to this section.