Review of Marza et al. “Towards the construction of regional marine radiocarbon calibration

curves: an unsupervised machine learning approach”

Marza et al., present an analysis in which they use clustering algorithms to investigate the spatial and temporal variability of marine radiocarbon reservoir ages in model simulations. The authors demonstrate that this type of analysis has a strong potential for improving marine radiocarbon calibration curves compared to the current standard approach of applying constant reservoir age corrections to the global marine calibration curve. I find the analysis interesting and agree there is a strong potential for this approach to lead to better radiocarbon-based age models for marine cores.

Comments:

Overall, my comments are minor. The problem and questions are well-introduced and the results are explained mostly coherently. My one complaint would be that I hoped the authors could go further in demonstrating how the cluster results can already inform marine core age models.

If this analysis were conducted on a transient simulation of the deglaciation, then the sediment-based R-age estimates shown in Fig. 14 could be used as validation of the k-medoids results, and the regional R-age curves from the cluster analysis would probably provide already provide useful estimates of reservoir ages for paleoceanographers. However, the interval selected from MIS 3 makes validation with proxy data difficult, and the regional R-age curves are probably less useful given large uncertainties in 14C ages from MIS3. Clearly, redoing the analysis on a new set of simulations is outside the scope of the current study, however it would be worth including some justification in the methods about why the two model datasets were selected and not a transient simulation of the deglaciation.

The brief section on the sediment-derived R-ages feels somewhat disconnected from the results of the models. I was hoping to see some validation that could show that the clusters found in the models are somehow reasonable when real datasets are considered. To that end – what does Fig. 14 look like if you apply the clusters from Fig 10 rather than calculating new clusters? Are the results similar to what you obtained from clustering the proxy timeseries? This would provide a good test of the cluster results and show their utility for interpreting paleo records of reservoir ages.



The authors suggest that applying these cluster methods to data-based reservoir ages could help develop regional reservoir age curves, and I agree. However, I am also wondering if it is possible to already gain some improvements by applying the results of the cluster analysis shown here without the need for generating new proxy R-age records. Please comment, are the k-medoids timeseries of R-ages already useful for constructing regional calibration curves, why or why not? Are there next steps that can be suggested to achieve regional calibration curves based on modelling/statistical methods?

Another idea that comes to mind for the application of these results could be a tool in which researchers can input a coordinate of a sampling location and then generate a map of the expected similarity of R-ages. With this information, one could select sites that have useful information about R-ages for your site of interest. Maybe an idea for future work…

Minor Comments

Fig 1 - Define U-Tr in caption. Over what timespan is this maximum variability calculated?

Line 240 – Line 245: I believe there are several discrepancies between how the text describes the figure, and what is shown in the figure and written in the caption. This makes it very difficult to follow. Some examples are below, but please check carefully.

Line 240 - Elbow is shown at K=6 in the figure, but written as K = 5 in the text.

Line 243 – K-medoids are on the right side of the plot, change “(Figure 4 left hand plots, dashed lines)” to say right hand plots

Line 244 -  “when using normalized data (Figure 4)” – should it say “unnormalized data”?

Line 245 – I am unable to parse the meaning of the concluding sentence of this paragraph.

Fig. 5 - Label time step units. Label A, B, C as noted in text.

Line 279 – Suggest deleting relatively.

Fig. 9&10 – label time step units

Fig. 11 – How was 7 chosen as the number of clusters? Based on the dendrogram, 4 clusters appears more logical to me (greater gap in the distances there), and would allow a more direct comparison with Fig 9.

Personally, I find Fig. S3 more interesting (comparing results from two different models and two different timescales of variability) than some of the plots included in the main text (eg Fig 12, Fig 6).

It would be interesting to also compare the K-medoids timeseries obtained from the U-Tr raw data with those shown in Fig. 9 and 10. I.e. what happens if you just use un-normalized data rather than doing a two step procedure with normalized data and then subclusters on raw data?

Line 417 – I think Figure 1b should rather be 1a

Thank you for letting me read your paper. I found it  really interesting.  Please find attached my comments - I hope they are useful, 

Tim Heaton