The paper "Simulation-Based Inference for Parameter Estimation of Complex Watershed Simulators" introduces a method utilizing SBI combined with deep learning techniques to improve the calibration of process-based simulators, focusing on Manning's coefficient and hydraulic conductivity in a snowmelt-dominated catchment. It aims to address two main challenges: the computational intractability of simulating complex watershed processes for parameter estimation and the uncertainty arising from simplified model representations of these complex processes. The study performed a series of synthetic experiments to investigate the performance of the SBI approach. The study is generally well-designed and the manuscript provides a detailed explanation of the methods, experiments, and results. Overall, I have no major concerns regarding the methodology of the paper. However, I believe that some points need further clarification. 

1) One of the basic assumptions of the presented research is the efficacy of LSTM networks as surrogates for process-based models. While the study acknowledges that LSTMs may not perfectly reproduce the behaviors, I think the LSTM model in this case study is still ideal (based on Table 3, the worst KGE can be as high as 0.77, which indicates generally good performance of the LSTM in mimicking the process-based model). This is understandable because the basin is dominated by snowmelt, which the LSTM does well due to its strong memory capacity. However, the broader applicability of the approach still needs to be discussed, especially given the known challenges LSTMs face in accurately representing hydrologic behavior in arid regions (e.g., https://doi.org/10.1029/2019WR026793).

2) My second concern is with the terminology used throughout the description of the experimental settings, particularly the interchangeable use of the terms "observation", "simulation", and "synthetic observation". I understand that observation refers more to the simulation of Parflow, but sometimes the distinction is not so clear, forcing me to rely on context to understand their intended meanings. This ambiguity is further complicated by the term "simulator", which is used variably to refer to both "ParFlow" and the "LSTM" model (see my specific comments).

3) A related point is: When I read the text discussing model misspecification, particularly lines 82-90, I initially interpreted the discussion as addressing uncertainties arising from the ParFlow model. This interpretation was influenced by the preceding context, which focuses primarily on the challenges associated with PB models. However, as I read through the manuscript, it became clear that the uncertainties referred to might actually be related to the LSTM model used in the SBI framework, since ensemble modeling is applied to LSTM models instead of ParFlow. Perhaps I misunderstood something, but it has indeed caused some confusion. If this is the case, the statement "calibrating these simulators to observed data remains a significant challenge due to several persistent issues including: .... uncertainty stemming from the choice of simplified model representations of complex natural hydrologic processes" in the Abstract would also be misleading,  as the manuscript primarily addresses uncertainties related to the LSTM surrogate simulator without adequately discussing the uncertainties inherent in the process-based (ParFlow) model itself. 

Specific comments are below.

Abstract: "watershed" and "catchment" are used interchangeably here and throughout the text (along with "basin"). I am not sure if the author wants to distinguish between them, otherwise please consider a consistent word.

l62-64: I think the two sentences are contradictory if "watershed prediction" is identical to "streamflow prediction". Perhaps watershed prediction here refers to the prediction of hydrologic states and fluxes other than streamflow, if so, please clarify.

l67: It would be helpful to state what the difference is between this study and Tsai's study in terms of how DL helps with parameter calibration of PB models.

l74: I would say that it is not always true that DL can "preserve fidelity" unless the surrogate has sufficient predictive power (see my major comment)

l88 (and Sec 2.4): Which model do you mean here, the parflow, the LSTM, or the neural density estimator? This is quite confusing.

Sec 2.2: It seems that SBI itself includes deep learning (l68), but here it looks like SBI can be independent with deep learning, please clarify.

l279: what prior distribution is used in the study? please clarify and justify.

l299-l304: which simulator do you mean here, I assume it is the surrogate simulator (LSTM)? Because you take the result of the parflow simulator as an observation. Please clarify and unify.

l316: Until I read here I am aware that you are saying the uncertainty from the surrogate simulator.

l380: How was the 14 days determined?

l391: It would be nice to show the NSE or KGE value here.

l391: I am curious about the applicability of the SBI approach to other basins in case the surrogate simulator does not capture the flow behavior well.

l431: Which simulator do you mean here?

l471: It is not clear to me what model structure you are referring to.

l542: I find it strange to call the LSTM result a "synthetic observation".

Figure 4: Since you have defined many "observations", please clarify which observation you mean here.

l559: please consider not using e-style scientific notation here (see inconsistency with l581 and l582).

Sec 4.1.1: may remove the header if there is no "sec 4.1.2".

l630: please justify why different settings of LSTM were used than in exp #2.

Dear Editor, dear Authors,

Please see my comments in the attachment.

Yours sincerely, Uwe Ehret