In this manuscript, Scapucci et al. investigate the impact of compound soil and atmospheric droughts on both ecosystem and forest floor carbon fluxes in a montane-mixed deciduous forest. Overall, the manuscript is well-written, and clearly, a lot of fieldwork and data analysis has been done, which is commendable. It has the potential to enrich the literature, and particularly, the idea of examining the responses of both above-canopy and below-canopy carbon fluxes separately to compound droughts is novel. I only have a few concerns that need to be addressed.

Major comments:

1) Since SHAP values could also give the overall feature importance, why not just use SHAP for the first driver analysis for daily mean NEP instead of using the conditional variable importance as presented in the manuscript? Are the results based on these two methods consistent? Air temperature is not important for daily mean NEP during CASD based on conditional variable importance in Figure 4, while air temperature is still important for daytime mean NEP during CSAD using SHAP in Figure 5. Although daily mean NEP and daytime mean NEP are different, the results using the two methods seem to be inconsistent. Therefore, please also report the overall feature importance of predicting daily mean NEP based on SHAP.  

2) The authors used the response curves of SHAP values vs. the abiotic factors to derive the driver thresholds. They observed an increase in Tair_NEPmax and a decrease in SWC_NEPmax from 2015 to 2022, thereby concluding drought acclimation of NEP to higher VPD and lower SWC. However, the thresholds they found were responding to the maximum positive marginal effects of drivers, which means the drivers at these thresholds increase the NEP. As drivers at the ‘real’ thresholds are expected to decrease the NEP during droughts, the identified thresholds are likely not relevant to drought. Instead, deriving these thresholds by taking values where SHAP values transition from positive to negative makes more sense. Furthermore, the SHAP values came from a single XGB model trained on the entire period of 2005-2022 (for NEP), which assumed that the NEP-meteorology relationship was stationary over the studied period. But if the drought acclimation of NEP exists, a shifted NEE-meteorology relationship is expected. Therefore, training the XGB models on each year separately (or several years using a moving window) and computing the SHAP values for each model might be a better option although the training dataset may be too small. Overall, the evidence of the acclimation of NEP to droughts is weak given the data availability and analysis, and thereby I suggest removing the associated results and conclusion if more convincing evidence is not found.

Line-by-line comments:

Line 22: What is the 30% decrease relative to?

‘largely’ → ‘large’.

Line 28: remove the second ‘always’; add ‘has’ after ‘net radiation’.

Line 31-32: remove the sentence of acclimation if more convincing evidence is not found.

Line 61: ‘be it’? ‘particular’ → ‘particularly’

Line 107: please add the description of measuring CO2 storage change.

Line 132: please report the depths.

Line 134: How to centerly normalzied the SWC data

Line 171: daytime mean NEP and daily mean NEP are easy to get confused in the many parts of the manuscript. Using ‘NEPdaytime’ and ‘NEPdaily’ could help.

Line 177: ‘Shapley, 1953’ is missing in the reference

Line 182: Please clarify why the mean SHAP value instead of the mean absolute SHAP value is used to indicate the overall feature importance

Line 189-195: please refer to Figure 7.

Line 212-217: All the events' length seem to be 1 day shorter. Same in table 1. Please check.

Line 237-204: What are those shade areas around dashed lines?

Line 241-243: Why Max. or Min. has a standard deviation?

Line 266-272: 1) What are those shaded areas around dashed lines in the left panels?

2) What are those error bars in the right panels?

Line 358-359: How to calculate this standard deviation?

Line 371-373:  Since SR vs. TS and SWC during CSAD are not significant, please rephrase ‘tend to decrease or increase’ as ‘non-significant’.

Line 385: If still keep ‘acclimation’, please briefly describe what acclimation is here.

Line 419-421: You found air temperature is not important for daily mean NEP during CASD based on conditional variable importance in Figure 4, while air temperature is still important for daytime mean NEP during CSAD based on SHAP in Figure 5. Although daily mean NEP and daytime mean NEP are different, the results using the two methods seem to be inconsistent. Therefore, please also report the overall feature importance of predicting daily mean NEP based on SHAP.  

Line 438-447: again, suggest removing if more convincing evidence is not found.

Line 492-493: same as above.

Authors research provides important insights into the response of a montane mixed deciduous forest in Switzerland to CSAD events, which are becoming increasingly prevalent due to climate change. By utilizing multi-year eddy-covariance CO2 flux data, authors have effectively characterized CSAD events at the study site and quantified their impact on ecosystem and forest floor CO2 fluxes. Authors used data-driven machine learning methods to discern the drivers of CO₂ fluxes which capture the complexity of these interactions. Overall, the study represents a significant advancement in our understanding of forest responses to CSAD events and highlights the importance of considering multiple drivers in predicting site-specific drought conditions and long-term forest responses. The only major limitation I see is the lack of information regarding the development of forest structure between the measured years 2015-2018-2022, but I believe that authors could address this easily. I am suggesting minor revision of the paper.

Comments:

Line 80: What does percentual cover mean for the species, by leaf area/volume?

The largest limitation of the presented study is bare minimal information regarding the forest structure. I believe that authors should include the development (annual) of standard parameters such as stand LAI and species specific DBH, height and density. This is especially important for the interpretation of the values between years and comparison with reference period. You should show that these differences were not due to differences in forest structure.

As this is managed site, the time between 2015-2022 is pretty long period that could include some significant change in species composition. This could influence your Figure7,8 comparison of variable sensitivity between years.

Line 82: First time mentioning Fraxinus excelsior, Acer pseudoplatanus etc. please use full latin nomenclature as you did for European beech and Norway spruce.

Line 151: R version missing.

I would suggest to include the variable of interest (NEP, Rff) in Figures 5-8 to include in the figures directly, not only in the description.

Could the figure 7c,f,I be interpreted in a way that the temperature optimum for NEP shifted between the years? If yes, I think you should explore possible reasons in the discussion.