This manuscript details a study that makes use of sap flow measurements in a catchment-scale model, through the canopy resistance parametrization.

This is an interesting topic since (i) sap flow measurements are not common, particularly at the catchment-scale and probably as a consequence, (ii) the current parametrizations of the transpiration process in hydrological models are poorly constrained.

The paper is well written, the objectives are clearly stated and the results are nicely presented. I have only minor comments concerning mainly the metrics used to assess the differences of the model outputs, the conclusion on machine learning modeling that is in my opinion over-optimistic, and some methodological details that are not detailed enough.

l.83-84: why not use the Machine Learning (ML) model on conductance estimates instead of sap flow data? Could this improve the ability of the ML model to reproduce gc_sap ?

l.160: what is the spatial resolution of the LAI estimates? How many tiles are considered in the catchment?

l.161-162: The term “reference model” for the simulation using the Jarvis-Stewart model could be changed since it may suggest that this simulation is the closest to reality.

l.188-190: There are missing details here. Why is it necessary to fill the gap in canopy conductance estimates? Why not just drop the concerned time steps? Why is it necessary to smooth the time series with a three-hour window? Since no details are provided, the reader cannot figure out if these choices were a priori or a posteriori choices. Anyway, a justification is needed here.

l.196: Similar to the previous comment, why use a sequence length of 96 hours preceding the prediction time step. Was this value optimized or chosen a priori?

l.225-227: I found the description of the results very incomplete. Canopy conductances estimated by sap flow and Jarvis-Stewart are also very different in terms of variability since gc_sap presents much higher temporal fluctuations. The discussion focuses on bias and Spearman correlation but I think that alternative metrics might be used to provide a complete figure of the differences. Please consider using the Pearson correlation coefficient and the ratio of variance, and/or the KGE.

Figure 1: This would be nice to add the streamflow time series since this is the only integrative measurement available. Also, a map showing the location of the available measurements of the catchment would help the reader to figure out whether the soil moisture probes are representative of the catchment. This could also be discussed in analyzing the results of Figure 1. I do not understand why transpiration rates are plotted only at the monthly time scale. Showing the high-frequency values would be valuable. Does the transpiration rate from sap flow are much more temporally variable compared to the simulated transpiration rate from the Jarvis-Stewart model? Is this why the conductance time series were smoothed by the 3-hour rolling mean?

l.247: “The Weierbach fell dry on 61 days (> 0.001 mm h-1) during the three-year record.” I did not understand this sentence.

l.264-265: I had an opposite interpretation of the outcome of adding 15 randomly picked days of the dry period. To me, this is proof of the lack of robustness of the ML model and proof of its inability to extrapolate.

l.310-320: In line with my previous comment, I found that the statements expressed in this paragraph are biased in favor of the ML model. I am not a great defender of complex and heavily parameterized models but in my opinion, ML models are also “complex and uncertain” and they suffer from overparametrization.

l.324-326: I think that this sentence should be placed in methodology to help understand why the ML model is used to estimate transpiration and not canopy conductance.

l.358: approach . -> approach.

The authors present a catchment-scale hybrid model which is leveraged by sap flow data for more accurate hydrological simulations. The results showed that the hybrid model could lead to more realistic soil moisture estimates than the conventional Jarvis-Stewart equation, especially during drought conditions. The hybrid model predictions could match soil moisture and transpiration equally well as model runs using observed sap flow data and more importantly, hybrid model has good potential extrapolation beyond the study site. Such kind of hybrid model approaches which integrate machine learning methods and physical laws could open promising perspectives for more parsimonious process parametrizations.

These very interesting results have great potential to benefit the scientific community. With some minor clarification, this manuscript will be considered for publication.

I just have several specific questions. First of all, they didn’t provide the cross-validation results. Secondly, did you also try the normal neural network, not the GRUs?

Line 129, 130: So, the 32 trees are evenly distributed in the catchment area? Could you show them on a map?

Line 196, 197: how many predictions time steps? Use 96 hours to predict next hour or next 2 hours? Why not 24h, 48h or 72h?

Line 197~198: How did you prove the network which consists of four layers (input, two hidden, output) with 128 cell/hidden state is the most appropriate structure? Will the different dropout rate affect the results significantly, e.g., 5%, 15%, 20%?

Line 260: So, you’re using data from 2014 and 2016 to train the deep learning model, while use the data of 2015 as the test dataset? Did you try cross validation and set 2014 or 2016 as the test dataset to see the results? Are there significant differences between different catchments and years? Could you show the data distribution, e.g., boxplot, of different years and catchments?

Section 2.2.4: It seems that the machine learning model is set to point to sap flow directly? Why not just let machine learning model predict the conductance directly? You could also introduce constrains into the loss function by using equation 1 and 2 to constrain the training process.

I also suggest you should have a flow chart or schematic map for clearly clarifying the hybrid model. This could be more friendly to the readers.

Line 293, 294: Could you further explain why the gcDL under- or overestimates on peaks? It seems that the model can’t not capture the peak value very well? I think if you let the machine learning model predict the conductance directly with constrains from equation 1 and 2 into the loss function, this problem could be mitigated.