The paper is well written and demonstrates a powerful modeling tool, especially with the capability of handling different rainfall intensities.  See comments on attached manuscript.

Review of “Quantifying fire effects on debris ow runout using a morphodynamic model and stochastic surrogates” by Spiller et al.

Summary

Spiller et al. present a simulation study designed to evaluate the ability of paired complex model-surrogate model to generate probabilistic postfire debris-flow runout hazard assessments both immediately after a fire and as recovery progresses. The manuscript is technically sound and the result are not over interpreted. The authors present an innovation that has the potential to substantially improve the process of postfire hazard assessment and I look forward to seeing the manuscript published in NHESS. My primary concerns are mostly suggestions for clarification or areas that I think more discussion is warranted.

Primary concerns

1. I suggest refining how the volume emulator introduced at L304 is described. As best as I can tell, the volume emulator takes either a subset or the full parameter vector (clarify) and produces only one value, the total eroded sediment (define s in L312). The volume emulator is then used as a filtering/screening step for the probabilistic analysis (described starting L313) to ensure consistency between the parameter sets and the Gartner et al. (2014) emergency assessment model.
2. The process for generating probabilistic hazard assessments (starting L313) assumes that some subset of parameters is conditioned on the Gartner et al. (2014) emergency assessment model and some subset of parameters is sampled randomly (state whether uniform distribution or something else). This seems like a reasonable first pass at generating a probability distribution at each map pixel; however, I also think there are a number of limitations of interpreting the results as a probability. For example, it assumes you have a decent prior distribution of the randomly sampled inputs. I suggest discussing the implications of these limitations, and potential ways forward, in the discussion section. What would it take to not condition on matching the Gartner model and what implications would meeting that bar have for portability (see point 5).
3. The results of Section 4.2 make me wonder how few simulations you can get away with. It would likely depend on where (in map view) you are located. Consider discussion how you might determine the minimum number consider the spatial variability of the runout area.
4. It is hard to interpret the results of the recovery experiment without seeing the curves of Liu et al. (2021). I think it would be worth reproducing those curves in this contribution, as well as stating the values for hydraulic roughness and saturated conductivity used for each panel in Figure 6. In the discussion you might also describe the potential to stretch curves based on satellite-based recovery metrics (Graber et al., 2023) or even use spatially distributed measures of vegetation recovery directly within the simulations.
5. The text at L403 describes the spatial pattern of erosion. This is very cool, and an important point to make to motivate future improvement and to highlight that the correct runout results can be obtained even if the upland recruitment is not quite right. However, I suspect it would benefit from a figure illustrating the finding.

I would be interested in the authors thoughts on the implications of this result for the portability of the emulators and the need for site-specific calibration. I could imagine that on the basis of this work, the authors might articulate a set of behaviors any emulator-based approach would need to demonstrate in order to be portable.  This list of candidate criteria could be highly useful for guiding future research.

1. I am also interested in the authors sense of how sensitive the result are to the shape of the design storm. I do not think additional work is needed to elaborate on this point. However, different storm shapes are commonly discussed in operational work (e.g., USDA SCS 24-hour storms, alternating block method). Consider elaborating on this point in the discussion.

Figure comments

• Consider showing the location of KTYD on Figure 1.
• I think the manuscript would benefit from a plot of the rainfall timeseries as well as the XX time series used to generate…
• Figures S1, S2, and S3 are referenced but not present.
• In Figure 3, I suggest adding a legend that labels the red circles as the left out points, and the blue/black as predicted. It took me a while to understand this figure because I was confused by the large range of the red values. It might be helpful to remind the reader in Section 4.2 that the training simulations are designed to generate a large range of flow depth values and a good test of emulator performance is the ability to confidently predict the left out value using the other values.
• I cannot easily see the flow edges in the figures that use satellite base maps (Figures 5, 6, 7, 8). Suggest revising.
• Suggest stating the other parameter values used for each panel in Figure 2.
• Suggest stating that no recovery is considered in the Figure 5 caption.

Suggested references

In the time since this was submitted, Dunne et al. (2025) was published. As it complements Alessio et al. (2021) and Morell et al. (2021) work, consider incorporating it.

Line level comments

L9 – Elements before ‘grain size’ are missing.

L10 – The contents of the sentence starting “Sensitivity analysis” seems to duplicate the content of the sentence starting on L9. Consider stating only once in the abstract.

L16 – Parentheses are missing around nearly all of the citations in the manuscript.

L31 – Clarify what is meant by ‘sources’. Do you mean debris-flow volume? As you show later, that is itself parameterized by other inputs (e.g., rainfall).

L39 – Consider saying that this is infiltration excess overland flow.

L67 – In this paragraph you might also point out that not relying on specification of a volume permits exploration of recovery.

L75 – Although not strictly necessary, consider consistently quantifying what ‘computationally intensive’ means throughout. E.g., core-hour per fine grid cell?

L105 – I suggest stating why San Ysidro and Oak creeks were selected over other runout paths in the region (Montecito Creek, Buena Vista Creek, Romero Creek).

L125 – Here and throughout, I suggest giving a name to McGuire et al. (2017)as implemented by Titan2D. E.g., Titan2D is not one model, but many and so in section 5.1 on model performance, it is a bit confusing discuss Titan2D performance. You could call it T2D-M2017 (or something better). Whatever you do, this is the section to introduce the terminology.

L134 – Define what is meant by ‘new material models’. I understand you are discussing different rheologies, or similar. However, I think most readers will benefit from a bit more explanation.

L135 – Because this manuscript is not about the computational speedup, I suggest the text describing details of implementation are extraneous.

L141 – Define U, F, and G.

L159 – Define c as the sum of c₁, c₂,…, c_k.

 L189 – Cite the source of this DEM.

L205 – Given that debris flows in general, and this debris flow, in particular, had a wide range of grain sizes, I suggest adding some explanation as to the interpretation of d.

L201 – State the AMR refinement criteria, levels, and any other specifications used.

L237 – W has another usage (introduced at L179). I suggest using a different symbol. g is also used (with subscripts) in equations (5), (6), and (7). Disambiguate the symbology. This equation also differs from that introduced by Heiser et al. (2017) in that beta and gamma should be subtracted from alpha.

I find it useful to point out that this misfit metric is commonly called other things after undergoing linear transformation (trimline ratio, threat score, critical success index).

L258 – Here and elsewhere: I found inconsistent use of tense. The run inputs were chosen. I suggest revising to consistently describe work done in the past in the past tense.

L271 – Earlier, the subscript k was for size classes. Consider using unique symbology for size classes and input parameters.

L270 – ‘Range parameter’ may be the standard name for Theta. However, because all the other parameters are inputs and the range parameter is a measure of sensitivity, I got confused. If it is not a standard name, consider using a more descriptive name. If it is a standard name, consider introducing it in a bit more detail (e.g., it appears like a variance such that low variance parameters are more important).

L274 – Here is a place where a consistent definition of computational efficiency would benefit. It would permit direct comparison between the complex model and the surrogate at a prediction pixel level.

L289 – The text in this paragraph does not seem like a numerical experiment to me. Rather, it is the analysis of the emulator fit.

L293 – State whether the emulator here is fit with N=64 or is something else.

L316 – This 10% range is larger than the uncertainty in the Gartner et al. (2014) emergency assessment model. Consider explaining in more detail why 10% was chosen and what it means in light of the larger uncertainty on that model.

Figure 4 caption – I₁₅ subscript formatting issue.

L433 – ‘with increasing delta’.

L477 – What would be required for a pre-fire emulator?

L505 – Consider providing an example driver file.

References

Alessio, P., Dunne, T., and Morell, K.: Post‐Wildfire Generation of Debris‐Flow Slurry by Rill Erosion on Colluvial Hillslopes, J. Geophys. Res. Earth Surf., 126, e2021JF006108, https://doi.org/10.1029/2021JF006108, 2021.

Dunne, T., Alessio, P., and Morell, K. D.: Recruitment and Dispersal of Post-Wildfire Debris Flows, J. Geophys. Res. Earth Surf., 130, e2025JF008325, https://doi.org/10.1029/2025JF008325, 2025.

Gartner, J. E., Cannon, S. H., and Santi, P. M.: Empirical models for predicting volumes of sediment deposited by debris flows and sediment-laden floods in the transverse ranges of southern California, Eng. Geol., 176, 45–56, https://doi.org/10.1016/j.enggeo.2014.04.008, 2014.

Graber, A. P., Thomas, M. A., and Kean, J. W.: How long do runoff‐generated debris‐flow hazards persist after wildfire?, Geophys. Res. Lett., 50, e2023GL105101, https://doi.org/10.1029/2023GL105101, 2023.

Heiser, M., Scheidl, C., and Kaitna, R.: Evaluation concepts to compare observed and simulated deposition areas of mass movements, Comput. Geosci., 21, 335–343, https://doi.org/10.1007/s10596-016-9609-9, 2017.

Liu, T., McGuire, L. A., Wei, H., Rengers, F. K., Gupta, H., Ji, L., and Goodrich, D. C.: The timing and magnitude of changes to Hortonian overland flow at the watershed scale during the post-fire recovery process, Hydrol. Process., 35, e14208, https://doi.org/10.1002/hyp.14208, 2021.

McGuire, L. A., Rengers, F. K., Kean, J. W., and Staley, D. M.: Debris flow initiation by runoff in a recently burned basin: Is grain-by-grain sediment bulking or en masse failure to blame?, Geophys. Res. Lett., 44, 7310–7319, https://doi.org/10.1002/2017GL074243, 2017.

Morell, K. D., Alessio, P., Dunne, T., and Keller, E.: Sediment Recruitment and Redistribution in Mountain Channel Networks by Post‐Wildfire Debris Flows, Geophys. Res. Lett., 48, e2021GL095549, https://doi.org/10.1029/2021GL095549, 2021.