The authors employed deep learning models to reconstruct paleomonsoon rainfall over India, using a range of paleoclimate records and two distinct approaches: one generating spatial map time series and the other producing regional time series. They implemented Convolutional Neural Networks for the spatial maps timeseries and multilayer perceptron models for the regional time series. The authors also incorporated an ensemble modelling strategy to enhance the robustness of their predictions. Although the predictive performance was not particularly strong, they effectively used the available dataset to perform data mining, and extracted as much information as possible, and I find that there methodological approach is pretty novel. Additionally, they applied explainable machine learning techniques to identify key predictors (paleoclimate records) across different locations, and supported their findings with physical knowledge.

The methods and results were well presented and thoroughly explained, with the authors fairly acknowledging the strengths and weaknesses of their findings and providing a well-rounded discussion. The datasets produced in the study have also been made available.

I recommend publishing this work, but I have a list of minor comments that could further improve it.

Specific comments

1. When using the term "timeline models", readers might expect an architecture designed to handle time dependencies, such as recurrent neural networks or transformers. However, your MLP model does not inherently handle sequential or time-dependent data. I suggest using "regional models" to refer to MLPs, as you are applying them to predict region-based time series, and "spatial models" for the CNN models that produce spatial maps.
2. Line 33: ENSO acronym not defined
3. Line 35: ITCZ acronym not defined, same for PDO in Line 59 and SST in Line 99
4. Line 84: ‘Machine learning approaches have been used in palaeoclimate research for automated palaeoenvironmental record generation, model post-processing’. Could you clarify what specific model is being referred to? Same applies to  ‘Machine learning has not been as widely used for model post-processing’ in Line 93. Please specify which models you are discussing to enhance clarity.
5. Line 98: There’s a typo in byMalmgren
6. Line 99 and elsewhere: The term ‘simple’ MLP is not commonly used. Consider using a more standard term such as ‘basic MLP’ or just ‘MLP’
7. L100: There is an extra ‘to’
8. Line 101: The description ‘two-layer MLP’ includes unnecessary detail. Consider simplifying it to just ‘MLP’
9. L116: Instead of ‘to stabilize model training’ it would be better to say: ‘optimize the performance of the model’
10. Figure 1: Are there blue contours present, or do you mean blue shades?
11. Section 2.1.4: This technical detail might be more suitable for the supplementary materials. Additionally, in point 6, please clarify the difference between the two CSV files.
12. Line 162: In ‘After this filtering, there were 157 datasets that could be used to train the models. (Fig. 1),’ please specify that these are the predictors of the models for clarity.
13. Line 203: how many years/samples were left for training?
14. Line 204: These are not ‘Independent models’ since they share training data. It would be better to say ‘separate models’ or ‘distinct models’ instead.
15. Section 3.2 Regularisation: Instead of writing this as a separate section, consider combining it with the loss function section and include the equation of the loss function with the regularisation term.
16. In Table 2, for clarity, update the caption to specify that L1 and L2 represent the hidden layers, and indicate the number of neurons in the input layer
17. Line 244: Was the choice of alpha arbitrary, or was it determined through trial and error?
18. Lines 255-258: The ‘Data Preparation’ section is unnecessary. Instead, you could include details on how the predictors were standardised or normalised, as this information is currently missing.
19. Figure 2: Where is the monsoon core zone?
20. Lines 285-290: Instead of detailing how CNNs are typically used as encoders, focus on explaining your specific approach; starting with dense layers and then using CNN as a decoder, as mentioned in Line 290.
21. Line 325: The term ‘distribution mean’ may not be accurate. It would be clearer to say ‘the mean of all predictions.’
22. In the caption of Figure 3: ‘Observed values taken from the reconstructed time series in Sontakke et al. (2008) are given in green’. Do you mean the observed values are shown to the right of the green line?
23. Line 367: The term ‘pattern correlation’ is ambiguous as it could refer to either ‘temporal’ or ‘spatial’ correlation. Please specify that this refers to ‘spatial’ correlation for clarity.
24. Figure 5 caption: Please revise the sentence ‘the prediction from with the PCC is computed is taken’ for clarity. Additionally, ensure consistency between ‘CNN-ERA5’ in the caption and ‘ERA5-CNN’ mentioned in Line 385.
25. Figure 7 caption: There is no dotted black line as mentioned; instead, there is a single dotted green line. Please correct the caption to reflect that it is one dotted green line, not ‘green lines’.
26. Line 455: Are you referring to low ‘temporal’ resolution?
27. Figure 8: How were the Shapley values standardised? Also, did you take the absolute values of the Shapley values before averaging them? Please add these details for clarity.

General Comments

In their manuscript Hunt and Harrison provide a novel approach to reconstructing historical monsoon variability: using machine learning to assess the relationship between paleoclimate records as predictors and the IMD gridded rainfall dataset between 1901-present as the target, and then extending that record back through time using annually resolved paleoclimate records from tree-rings, speleothems, glaciers etc. The results are interesting and appear to provide a significant step forwards monsoon predictability, comparable or slightly better than previous techniques in developing time series at point locations or areal means, but a significant advance in reconstructing spatial heterogeneity of monsoonal variability. The section on assessing the importance of individual paleoclimate records in the model, and the ‘sphere of influence’ of individual sites is an exciting development and appears to be backed by our understanding of monsoon dynamics.

The manuscript is well written, and the language is largely appropriate but could be simplified in places to the intended audience of Climate of The Past who likely have a more limited understanding of machine learning techniques. This includes more high-level introduction to methods, and care explaining predictors, targets, training datasets, validation datasets, test datasets etc. There is also sometimes a disconnect between what is labelled on the figures, the figure captions and what the figures refer to. Consistency and clarity are needed in places.

I do not have any significant concerns over the science and resulting inferences presented in this manuscript, though I will happily defer to more qualified machine learning reviewers on the robustness of the methods. I believe that this paper is suitable for publication in Climate in the Past, after correction for some minor issues.

Thank-you

Nick Scroxton

Maynooth University

Specific Comments

• Could you explain LiPDs in either section 2.1.1 or 2.1.2 before you get to line 148.
• Section 3: A high-level couple of sentences at the start of the methods would help non-machine learning experts. For example, it doesn’t actually say in the manuscript what is your predictor data-set is and what is the target data-set.
• What’s the difference between validation and test datasets in Figure 2 and section 3.5 1a. What do they do and why?
• Figure 3: This figure does not show clearly the information that is attributed to it. There are significant mismatches between the figure and figure caption. I don’t see the ensemble median in black, the spread in red, or the Sontakke time series in green. I think this figure should be expanded to take-up more space on the page to really highlight the key results attributed to it.
• Line 380-383: The dismissal of the PCA technique is too strong at this stage of the manuscript. At this point the reader has only been introduced to figure 4. The PCA outperforms the CNN model 40% of the time (2 out of 5 years) in figure 4 so cannot be dismissed, although an argument can be made that it lacks spatial heterogeneity. Once we have seen figure 5 and the larger dataset, we see that outperformance of the PCA method in figure 4 is likely just an artifact of small sample size, and therefore the dismissal is more reasonable. This section therefore might need to be reworded
• I wonder if dry years also correspond to major volcanic eruptions (as we might expect). This might provide an additional test of reconstruction performance that is less reliant on additional human societal complexities.
• I disagree with some of your speleothem inferences. In section 4.3 If the EPI speleothem record is d18O then we might expect a better correlation with WPI rainfall than EPI, and thus this example belongs in the following paragraph instead. I don’t see the relevance of the river basins argument here on speleothem proxy variability.

Technical Corrections

• Line 196-198: Could you unreversed the first half of this sentence for clarity
• Figure 2 caption: ‘one of the timeline model’
• Line 354: New paragraph?
• Line 364: 1988?
• Line 380: New paragraph?
• Line 528, 542: The cave name is spelt ‘Mawmluh’ or ‘Krem Mawmluh” if cave needs to be included.
• Line 530: ‘documented’?
• Line 531: ‘subcontinent’ to be consistent throughout the manuscript.
• Line 561: clarify what you mean by wavelength or use less technical language.
• Line 578: delete ‘have’