This is my review of "Multi-fidelity model assessment of climate change impacts on river water temperatures, thermal extremes and potential effects on cold water fish in Switzerland” by Love Raman Vinna et al., submitted to Hydrology and Earth System Sciences. In this paper, the authors combine an ensemble of climate models, hydrological models and two water temperature models with varying parameter settings per station to derive water temperature projections in Swiss rivers. In addition,  they determined future changes in projected extremes, hysteresis effects and impacts on brown trout. I found the study interesting, with sound methodology. However, the manuscript would benefit from a clearer presentation to fully convey its strengths. I therefore suggest minor revisions, and I provide comments below to help improve these aspects.

Major comments:

While comprehensive and detailed, the data and methods section is somewhat difficult to follow due to its length and repetition. I recommend a thorough revision to improve its structure and conciseness, eliminating redundant wording. This will enhance readability and streamline the section. Below, I provide specific examples and suggestions to support this refinement.

The presentation of results could be refined to enhance clarity and readability. Figures 4, 6, and 7, along with Table 3, contain a wealth of information, but an additional or alternative figure presenting the data in a more aggregated way—such as by thermal regime—could help highlight key differences more effectively while still accounting for uncertainty. While the station-based results provide valuable detail, incorporating more synthesized figures or tables could make the main findings more accessible. Additionally, summarizing key insights more narratively, rather than listing numbers extensively, may improve the flow of the results section.

Specific comments:

Title: the “cold water fish” might suggest a more elaborate analysis on general fish species when only brown trout is considered in the analysis. Suggestion to rephrase/remove.

L 16-18, abstract: Provide more detailed “Alpine thermal regime” and “Downstream lake regime”.

The introduction could benefit from a better gap description, and background building up to this gap description, for example highlighting the gaps of current water temperature projections for Switzerland available in literature. Also, since part of the novelty of the study lies in the modelling approach employed, this gap can also be made more apparent.

Figure 1: Good overview figure, although it would benefit from a distinction between data and data sources and operations on that data in the flow chart. This is an open suggestion

L73-92: this is a very general explanation about the choice of models to use, I would suggest to move it to the intro

L126-138: can you give a little more extensive description of the CH2018 climate data? E.g. introduce that they are derived from the EURO-CORDEX regional climate modelling ensemble, with the number of RCMs driven by GCMs and future scenarios, as well as the horizontal resolution. These names return then in Table 1, allowing the reader to understand where they come from.

L130- ...It would help to also introduce the Hydro-CH2018 in a little bit more detail afterwards, with a subclause per hydrological model on its characteristics (semi-distributed, empirical etc). A suggestion is to structure the description of the different data sources with different subtitles

L151: I might have missed it, but how many monitoring stations are eventually used in the study? If these are the stations on Fig. 1 panel a, provide a reference to that fig (also for its other panels).

L167: DIS criterion: what is the threshold used to assume horizontal distance is “minimal”?

L168: how is the representativeness of the meteorological stations to upstream drainage area assessed?

Table 2: I would suggest to move this table to the appendix, as it is very extensive and does not add much to the results.

L179: “were already statistically downscaled”, please add details on how this is done. This could be part of the paragraph where the CH2018 scenarios are explained in more detail. Also, how big is the bias, and how much would it impact the results?

L222-228, suggestion to move this paragraph to after the model description.

L248-250: The fact that for each river monitoring station, the best water temperature model is employed is a key strength of the study, and should, to my opinion, be more pronounced throughout the study (eg intro describing the gap on this, abstract and earlier in the methods where the multi-fidelity is mentioned first).

L287-320: there is some repetition in this section, and parts are more difficult to follow, please consider condensing it retaining the same information

L332: could you give a more explicit explanation of a hysteresis example here, relevant for Swiss rivers?

L357-361: this is a good example of lines that could be shortened.

L368-372: if an extreme is defined by the deviation of the 90th percentile compared to the median of a certain 30 year period, why does this period need to be detrended? i.e., there is no certain time (beginning or end of period) where the analysis is carried out? Or am I missing something?

L396-400: If this info would be presented in a small table, it would be easier to grasp

Section 3.2 on hysteresis analysis. For a non-expert in hysteresis, I found the results difficult to interpret. To my opinion, it would be beneficial to provide some guiding sentences on how these results could be interpreted.  

Section 3.3 and figure 6. It should be more clear from the start of the paragraph that the “extreme event severity index” is used, so the values do not represent absolute extremes, but deviations from the “normal” in the respective period. L477-479 indicate this, but it should be more up front in the paragraph and figure to avoid confusion for the reader.

L515-525: these are very general sentences which fit better in an introduction section than in a conclusion

L525: why are the water temperature models of “lower fidelity”?

L587-593: why is it important to study these hysteresis effects?

L605-614: I would move these lines to the results section, and provide more explanations on the differences between thermal regimes here in the discussion.

L634-652: same comment as above, these lines are more for the results section, which would make that section more digestible.

Textual comments

L93-94: “multi-fidelity modelling” and “from multiple different fidelity levels”, I would avoid such repetitions in the same sentence

L111-114: this is repetition of what is said above

Caption Table 1: “hydraulic models” or “hydrological models”?

L161: “Only stations...”, hydrological measuring stations are meant here, I suppose?

L170-172: “For situations ...”, You lost me in this sentence. Would it be possible to reformulate more clearly?

L374: section title: indicate that the thermal thresholds are for fish.

L421-422: reformulate “at for each station for with ...”

L667: suggestion to just name this section “5. Conclusions”

Råman Vinna and colleagues have used a coupled climate-hydrological-temperature model setup with different levels of representation of reality (fidelity) to simulate river temperatures in Swiss rivers, including future climate projections. They include extreme temperatures and ecological thresholds in their analysis. The topic and scope are rather similar to the Michel et al. (2022, cited in text) paper, also published in HESS, but they expand on it by their multi-model approach and additional analysis of thresholds, extremes, and hysteresis. The paper is well-written, clear, and comprehensive, and I did not have major comments, only some minor and mostly technical ones that I outline below.

Minor and technical comments:

• 24 -> “but also in rivers where…”
• 239: -> “inverse stratification”
• 261: I cannot find Table C2. I assume you meant B2?
• 332: “refiling” or “refilling”?
• Figure 3: Change y-axis to “Water temperature”
• 357: -> “straightforward”
• 362-373: I like this definition of a severity index
• 421: “at” or “for”
• 450: -> “from, for example, …”
• The caption of Table 3 mentions yellow marking which does not occur in the table.
• 525-528: It is unclear from this sentence what vital principle is referred to.
• 552-556: Are there 15 or 16 Alpine stations?
• 585: -> “Rhône”
• 584-586: Could you rewrite this sentence? It is not clear whether lakes or rivers warmed faster in the cited reference.
• 625: “brown trout’s” -> “brown trout”
• 675-676: 0.37 °C/decade over 11 decades would be a 4 °C increase. Could you double check the numbers? Moreover, the Results mention a 3.18 °C increase total and 0.36 °C/decade, so please standardise these values.
• Discussion: One major process for river temperature in the studied systems seems to be the role of disappearing glaciers and snow cover. This is also likely an important factor for changing hysteresis patterns. The discharge models used in the paper may take this into account, but air2water/air2stream do not (e.g. L. 280). It may be valid to assume a nonlinear response of river temperature to disappearing snow/ice that may not be reflected well in the training data (especially for Alpine streams and their low number of stations with long measurement time series). It would be good to add a short paragraph to the Discussion on how this may affect the projections in this paper.

This paper introduces modeling climate change impacts to river temperatures in Switzerland. To do this, the authors conducted a multi fidelity modelling method which uses statistical pattern recognition to estimate river water temperatures under climate change and thereby close the aforementioned spatial gap by determining, in an automated manner and on a country-wide scale, how future river water temperatures are likely going to change.

The authors frequently refer to their method as novel. I suggest to remove all occurrences of this claim of novelty. Simply describe the model used. Some would argue that the discipline of stream temperature modelling has advanced beyond the use of air temperature and discharge alone for predicting river temperature, regardless of whether it is focused on current versus future climate. While it may be practical for a nationwide attempt, and in that case also ‘efficient’, it is not necessarily “novel”.

The reason physically-based models require a lot of data is because they attempt to represent mechanisms and therefore attribute causality of rising river temperatures. River temperatures are a function of many processes beyond simply river discharge and air temperature, as has been discussed in recent literature. The limitation of the “efficient” model approach is that many, many physical drivers of river warming are completely ignored. In predictions of stream temperature, simplifying the “more complex processes into purely empirical parameters” often involves using lumped parameters and lumped heat exchange coefficients which ignore aspects of climate change, especially with respect to the shortwave and longwave radiation balance, and increases in atmospheric emissivity which is driving the air temperature warming. There is not a single mention of any of this. The authors simplify the controls to the energy balance as being based on discharge and air temperature, which is not complete, nor does it use best-available-science. If the authors make simplifications in processes, and vary the number of parameters used across their different simulations in order to get a nation-wide dataset for Switzerland, they need to be very clear about this approach and also be upfront about the many, many limitations of their results.

Lines 122-125: It is unclear how many years of actual data were used. This must be clarified. In one sentence, they say at least 1 year, in another sentence they say “data should preferably cover 30 years”. Authors need to specify which simulations used which timespan of datasets, as this is a fundamental influence on the accuracy of the predictions you are reporting in your Results section.

Lines 151-153: The authors state, “For monitoring stations at which historic river discharge data or future river discharge projections weren't available, only future near-surface air temperature projections were used to simulate water temperature.” This is a major limitation. For how many stations did the authors predict river temperature only from air temperature alone? And how do you correct for the fact that come used discharge and some didn’t use discharge, but you are presenting the results of those two different simulation approaches as being equal in your Results section?

Lines 154-156: Many studies have demonstrated that the resolution of the climate model data will influence your results. Here the authors state, “Where climate projections were available at multiple different spatial resolutions (i.e. 0.11° and 0.44°), only one model, as indicated in Table 1, was included in the analysis, following the approach of Muelchi et al., 2021.”

These two items above both will affect the model results, potentially significantly. Sometimes the authors use air temperature and discharge to predict river temperature. Sometimes the authors use only air temperature to predict river temperature (many authors have shown this is not sufficient). Sometimes the authors used 0.11° spatial resolution and sometimes they used 0.44° resolution. How are the results defensible and comparable?

Lines 165-172: Again, the deviation across methods raise concerns for presenting comparable results. This study employs large datasets which require some level of computational proficiency, but it appears they did not employ spatial interpolation methods of weather data across elevation or across distance. It is very common (and not difficult) to employ spatial interpolation methods of time-series weather data to a particular river location, in order to produce more accurate results at a specific distance along a river. The authors state: “Meteorological stations were subsequently paired with hydrological stations such that (a) the horizontal distance between river and meteorological stations was minimal (criterion "DIS"), (b) the meteorological station was representative of the conditions in the upstream drainage area (criterion "DRA"), and (c) the elevation difference didn't exceed a reasonable threshold of 200 m (criterion "ELE"). Where possible, all three criteria were adhered to. For situations where the closest meteorological station was either not fulfilling DRA or ELE, the DIS criterion was evaluated only for stations which fulfilled both DRA and ELE.” While this explanation is, in theory, reproducible, I am not sure that adjusting the criteria on a station-by-station basis is defensible. Authors need to address this.

Lines 322-324: What do the authors mean by “shape-preserving interpolation” across multiple days without data, and where is this interpolation method presented in this paper? Authors state: “Before adjusting the water temperature model output from 1990 to 2099, Bcs was combined into a continuous dataset by filling in the 3- to 5-day gap in between each season with shape-preserving interpolation.”

Line 439: “Considering only the far future”  what do the authors mean by “far future”. Please clarify.

The authors’ most significant result is summarized by “Climate change impact was heterogeneous between stations, yet common patterns were found within thermal regimes”. It is concerning to present results when each result was achieved through a subtle deviation from the methods, the spatial resolution of inputs, the handling of missing days of data, and even using different model inputs. In some simulations the only model input is air temperature.

How can results and hysteresis loops be viewed as comparable across simulations by the reader, when the methods employed to get there were modified, changed, required deviation of some methods, used a different number of parameters in ‘air2water’/’air2stream’ (i.e. Line 695 “adapting their parametrization complexity to the required level”), or were slightly different methods across simulations?