General comments

In this paper Ensemble Echo State Networks (EESNs) are built based on data from Earth System Models (ESMs) to investigate the machine learning explainability technique of spatio-temporal zeroed feature importance (stZFI) to see the effect of different variables on each other in case of a large stratosphere aerosol injection, in this paper the natural event of the volcanic eruption of Mount Pinatubo. The paper contains several experiments ranging from simpler models to reanalysis data, showing the potential of the method. Although the main idea of the paper is clear, the details can be hard to follow, partly due to lack of explanation, or by using not the correct term.

Specific comments:

Line 44-45: What is a replicate hold out set vs a repeated hold out set? I think some more explanation would be good.

Line 88: How do these temperature differences propagate over time?

Line 94: When you are using reanalysis data, you are looking at a combination of models and observations, how would you get interesting relationships in only observations from this?

Line 206: Replace ‘real’ by ‘reanalysis’

Line 242-243: I am not sure I understand the difference between this paper and McClernon et al. (2024), would it not be better to also take T050 into account to forecast T1000? 

Line 246: Ensemble -> Ensemble member. The difference between ensemble and ensemble member is not clear throughout the paper.

Line 255-261: To be more clear, I think it would be good to guide the reader a bit more through Figure 1, why the temperatures in the NH have a negative anomaly for example, why the aerosol spread is much faster over the NH etc.

Line 261: Could you explain why there is this unrelated spike?

Line 267: How much time-lag is there?

Line 269: I do not really see smoother decay for T050?

Figure 4: Why is there a negative value for importance after the positive values for T050?

Line 326: I think it would be good to explain a bit more about Figure 5, what we are seeing and why. T2m for instance has a large positive anomaly around 1996, this is an outlier?

Line 347: Looking at Figure 6, there is no anomaly for T2m around 1996. Does Ensemble member 1 have that much influence on the feature importance?

Line 375: Figure 8 is about MERRA, not about E3SM?

Line 377: There is again a large anomaly for T2m, this time around year 1998. Could you explain this further?

Line 449: Could you add some explanation about whether this increasing value for T2M is significant?

Line 518-526: From the text is not clear which figures are for which models/reanalysis. The text about Figure B2 is equal to the text about Figure B3.

Technical

Line 86: Temperatures -> temperatures

Line 125: without -> Without

General comments:

Ries et al. (2024) primarily apply the spatio-temporal zeroed feature importance (stZFI), an explainable AI (XAI) tool, to investigate the relationships between various variables associated with a stratospheric aerosol injection event. Notably, this stZFI method can reveal how the feature importance of predictors evolves over time. Utilizing this approach, the authors evaluate the time-variant contributions of volcanic aerosols to the prediction of local and surface temperature. They validate the results with mutiple datasets, including both model simulations and observations, demonstrating that stZFI can identify relationships consistently accross different datasets. This article showcases the capability of stZFI as an exploratory data analysis tool in climate research with great detail and precision. However,  I would recommend the authors to devote more effort to explain the stZFI results physically. Please find my comments as below.

Specific comments:

Line 9-10: The meaning of this sentence is unclear to me. Here the authors only use feature importance to distinguish between signal related to volcanic aerosols and others, not really natural climate variability.

Section 1.1:  It is necessary to include the possible latitudinal transport of volcanic aerosols driven by the large-scale circulation in stratosphere - the Brewer-Dobson circulation (Butchart 2014).

Figure 1:  The movement of aerosols from equator to polar regions could also be driven by the Brewer-Dobson circulation, not only just due to diffusion. Please clarify it.

Section 2.2.2: Why do you only consider latitude bands for regional contributions? Is there meridional transport of volcanic aerosols? If yes, it would be interesting to give a latitude-longitude global plot showing regional feature importance when T050/T2M peaks.

Line 233: What’s the highest height of model outputs?  

Figure 4: What does negative importance in three subfigures mean? When can people trust that the feature importance from stZFI is reflecting a real relationship?

Line 439-440, Line 448-449:  The T2M FI shows a large increase over 1997/98 (Figure 10, subfigure for T2M). Could the increase of FI in this period be caused by the internal variability instead of the volcanic aerosol radiative effect? For example, there is a strong El Niño event from May 1997 to May 1998 (Wang and Weisberg 2000), which could lead to higher autocorrelation in T2M.

References:

Butchart, Neal. "The Brewer‐Dobson circulation." Reviews of geophysics 52.2 (2014): 157-184.

Wang, C., and R. H. Weisberg, 2000: The 1997–98 El Niño Evolution Relative to Previous El Niño Events. J. Climate, 13, 488–501, https://doi.org/10.1175/1520-0442(2000)013<0488:TENOER>2.0.CO;2.