Overall: In my view, this paper is what it says it is, i.e.: “a useful tool for early warning systems and current and future coastal flood risk analysis” for a broad range of fringing-reef morphologies. Given the high uncertainties of coastal flood risk and lack of available EWS across much of the world’s vulnerable reef-lined coasts, it makes it very worth reporting.

However, the paper could be substantially improved in several ways. Among them:

The authors should be more up-front and clearer about what is different between this paper and the earlier BEWARE paper (Pearson et al., 2017), which many readers may already be familiar with.  Besides the addition of (a lot) of new (real-world) profiles and related training data for the surrogate model, why has the Bayesian approach apparently been abandoned? Or are you just not calling the training steps “Bayesian” anymore?

Related - most of the complex logic appears to be used for matching the target reef profile to the representative reef profiles (RRPs, which were developed primarily in an earlier work); comparatively simple inverse distance weighted interpolation of the full-fidelity (XB-NH) model “training data” is then seems to be used to estimate target profile and target conditions (albeit with some interesting heuristic relationships used to post-hoc estimate effects of bed friction and beach slope).  That seems (in my experience anyway) a different approach compared to most coastal hybrid/meta-models, which seek to emulate the dynamics themselves over a given morphology (e.g. Zornoza-Aguado, et al 2024).  Would it not be easier (in the modern age) supply all training data (including the reef profiles themselves) to some kind of conditioned neural network (NN), either a simple one, such as the RBF approach used by Rueda et al 2019 and others, or a deep NN, or explore any of the rapidly evolving more complex black-box ML approaches? Maybe you don’t need that level of complexity due to the profile 1-D nature of the problem and a more first-principle morphological approach is better? I think the explaining the rational used here and how it diverges (or doesn’t) from other contemporary meta/hybrid modelling approaches for coastal extremes would greatly improve the paper.

The validation presented is limited to comparisons between the full-fidelity (XB-NH) model and the surrogate model.  While this is the norm for many hybrid modelling studies, it would be nice to see some comparisons of the surrogate model (alongside XB-NH) to real-world observation as was done in the earlier BEWARE paper. There are lots of empirical/statistical/analytic/hybrid approaches that estimate wave runup – how much better is BEWARE-2? Given the information, it is difficult to assess how much better BEWARE-2 might be compared to these other approaches.

Abstract:

Are the unit details on verification necessary? The upper limit runup of range (20.9 m) is non-intuitive until the semi-infinite beach slope is defined in the methods section. In my view it would be better to normalise RMSE and bias and perhaps represent them as percentages for the abstract so this stated range is not needed.

A little difficult to follow… also, what is the difference with this paper and the earlier BEWARE paper (https://doi.org/10.1002/2017JC013204)? That is front of mind to readers such as myself, who are aware of the earlier work.

Introduction

Ln 30 - : since publication of Hoeke et al 2013, the number of case studies attributing remotely generated swell as the primary proximal factor in island flooding events has expanded – I recommend adding a few more recent examples (e.g. Wadey, et al 2017, Ford et al 2018, Wandres, et al 2020, Hoeke, et al 2021) to highlight its pervasiveness among oceanic islands.

Ln  64: (Pearson et al., 2017; Rueda et al., 2019; Liu et al., 2023), consider adding  Beetham and Kench, 2018 to this list? Also, while all of these meta-modelling approaches may suffer “limited number of schematic coral reef bathymetries” how do their approaches compare to BEWARE-2? Is BEWARE-2 only better because more training data has been introduced or are there other improvements/considerations in the overall approach?

Methods

Ln 94-115: I found this section circuitous and hard to follow, with poor economy of words. At the very least end Ln 98 with “… using morphological clustering technique, as summarised in the following paragraph.”

Figure 2: This just looks like random coloured spaghetti – maybe sorting by mean profile steepness or runup would make this more sensible? Also, runup based on what boundary conditions? Is this normalised somehow?

Ln 299 “… 5, 25, 50, 75, and 95% depth exceedance values, i.e., the depth exceeded by a given percentage of the observed profiles at each cross-shore location” not sure I understand this …

Benefits and limitations and/or Conclusion sections:

I think it would be worthwhile to point out that the reef-lined coasts of many nations do not have the high resolution bathytopo information (e.g. based on LIDAR surveys) needed to make use of tools like BEWARE-2 – this paper is opportunity to point out the extremely high value of such underpinning data.

General comments:

This paper is concerned with the development and application of a meta-process modelling system to address the need for a fast, robust prediction of runup on reef-lined coasts. The scientific significance of the paper is substantial given that it addresses a very real problem associated with the need to better predict coastal flooding along reef-lined coasts. The use of large data sets which are validated against the results of a numerical model (XB-NH), and the incorporation of roughness variations make this an important contribution. The paper is generally well written and well presented. Progress towards an early warning system for such vulnerable areas would be highly beneficial.

Specific comments:

Though the paper is very thorough and uses large sets of data, as the authors suggest, there is a skewed focus on U.S. data, and testing the model with examples from other locations would be interesting to see.

The validation against the XB-NH runup values is understandable given the complexity involved with obtaining field measurements, however it would be interesting to see a comparison with field data, even if only for a very limited number of scenarios. The use of a 1D model is certainly far more practical, however would validation against a small set of scenarios with field or physical model data help to reduce the uncertainty as to the extent of these effects on the runup values?

I agree with the comments made by RC1, that a slightly more detailed/clearer explanation of the differences in the methods used in BEWARE2 compared to BEWARE would be beneficial to the paper.

Figure 2. I think this figure could be improved. I understand the use of the 195 RRPs but is there a deliberate order to the way they are presented? Could this be improved? Is yellow the best choice of colour for the low runup?

Could the font size be increased in Figure 3 and 5?

Figure 4: Is it necessary to include all profiles? The grey can barely be seen when printed.

I may have missed this, but what were the computer specifications used to run the XB-NH simulations?

Technical corrections:

Ln 191 “…converted on…” should this be “…converted into…” ?

Ln 472 “…”the influence reef health…” should this be “…the influence of reef health…”?

Ln 474 “100s years…” should this be “…100s of years…”?