The study is heavily based on the Lagrangian simulations of particle sinking based on circulation model results and described in Wang et al., 2022 (I think this should be stated more explicitly and include a brief summary of results. Best in Section 2). The authors use the results of these simulations to test the possibilities of using machine learning to locate the origin of particles caught in deep sea sediment traps. The aim is to use observed sea surface conditions (mainly satellite altimetry and SST) to deduce the source of particles.

The authors demonstrate that the machine learning algorithm vastly outperforms the baseline prediction, which uses the average distribution centred above the sediment trap. However, they should give a more detailed explanation of what are the benefits of using their algorithm over running the Lagrangian simulations. To put it in a anther way, why not simulate the particle paths using existing ocean reanalyses instead of ‘guessing’ the paths from surface observations? There are many regional model reanalyses of ocean currents available and e.g. CMEMS offers global reanalysis in roughly 8 km resolution which matches the one used in this study. It is clear that there are computational benefits, but one would think that with all the efforts invested in setting the sediment traps, dedicating some serious computing time shouldn’t be a problem. The answer might be obvious to the authors, but certainly not to the potential audience. The authors state in the introduction that there are uncertainties stemming from the ocean model errors. That is certainly true, but in such case I am missing at least a rough comparison of the accuracy of their method to the uncertainties in the prediction of the catchment areas using the Lagrangian backtracking.

I am not a native speaker, so my comments on the English grammar are merely suggestions.

Comments by lines:

L. 7: “surface conditions appear to be sufficient to accurately predict the source area” – Are they? What is considered sufficient?



L. 66: Why 50 m/day?



L. 76: “After spin-up, the chaotic evolution results in uncorrelated dynamics” – How are the initial and boundary conditions perturbed? Shouldn’t the same atmospheric conditions ensure similarity of both runs? Large distance from the boundary and same atmospheric forcing would imply very similar oceanographic conditions. This is very important. If the test conditions are related to the training conditions, the performance of the ML algorithm is overrated. This could also (at least partially) explain the results which are better “under weak/or stable dynamics”. Such conditions are likely also the situations when POLGYR1 and POLGYR2 currents would be most similar. This should be checked and thoroughly explained in the text.



L. 90: Why 10 km x 10 km patch? This paragraph is a bit confusing. 36 particles are released from 10 km x 10 km patches, but there are 36 STs in the model and they are 36 km apart. Is this right? So 36x36=1296 particles are released every 12 hours? Why not release them from 36 points instead of patches? To compensate for the lack of dispersion? The paragraph should be rewritten and clearer.

L. 104: I think the delta sign is not a good choice for the number of points as it implies the distance between points.

L. 108: The authors should explain better the 8 km resolution. I think that the simulations run at 2 km resolution, but the end results are downscaled and so are the surface fields used for the machine learning. This would also match roughly with the resolution of the satellite altimetry. Am I right? If so, this should be clearer from the text.

Figure 3 caption: remove “the” from “the origins”. “from the left panel” instead of “of the left panel.”

L. 142 (something wrong with the line numbering here): “the training criterion”. Batthacharyya coefficient is used in the equation as BC and should be marked as such.

The sum (epsilon) in the equation is too small and is missing the summation index.

L. 145: “Empirically, this method improves the performance of the trained model compared to experiments where the training loss is based only on BL_200m.” - This is very interesting. On one hand, one would expect only the end location to matter, on the other hand, the path is important. It would be interesting to know more.

L. 160: The values of BL seem a bit arbitrary. A visual analysis is kind of a weak argument. What does it mean in the practical sense? Is it possible to relate how much would a certain value of BL affect the content of the sediment trap? L. 184: “analysis” instead of “analyse”.

L. 189: “16 in)” – the “in” is kind of out of place here.

L. 219: Why the 20^th day? Shouldn’t these be averaged over the average travel time?

L. 256: Maybe “collected” instead of “measured”.

L. 275-285: The biogeochemical model results and satellite chlorophyll concentration measurements show the spatial variability of phytoplankton biomass. Maybe this could serve as a measure of needed accuracy of the method? On the other hand, this paragraph focuses on primary production only and neglects other sources of particles such as zooplankton. The latter could be obtained from the biogeochemical model as well.

L. 298: “analysis” instead of “analyse”.

general comments

This paper presents a novel method for estimating the near surface origin of sinking particles that reach a sediment trap at 1000m. A neural network based on the U-Net architecture is trained to predict backward trajectories of sinking particles from environmental variables. The authors present the method and demonstrate its skill in a twin experiment setting. They also provide compelling analysis of the behavior of their estimation method depending on dynamical situation. 

I feel this is a strong paper and recommend it to be accepted for publication after minor revision. The paper is generally very well written and clear. The method being introduced opens up interesting perspectives to help better interpret sediment trap observations, but the authors also acknowledge difficulties that lie ahead if such method was to be applied to real data. One of the strengths of this paper in my view is the analysis of results in relation to physical processes. A weakness is that we don't know how/if the method's skill could be quantified in the real world.

specific comments

- L66 : please provide a reference to justify the choice of "50m/day" as realistic on average in the real world, assuming that this indeed is the case. Is 50m/day representative of a large fraction of sinking biomass that sediment traps capture? Wang et al 2022, cited later, appears to be a modeling study by the same authors. Please state in the paper whether the choice of 50m/day is backed by real world observations, or otherwise.

- L107 : "based on the statistical results of Wang et al 2022" is vague. What statistical results?

- L138 : please explain and/or provide a reference for "Combining these steps with skip connections enables the detection of hydrodynamic structures at different spatial scales". 

- Fig 4 : more explanation about "concatenation" and "channels" would also be useful in section 5.2 for the non specialist reader.

- L178 : can it be excluded that Unet4layers simlply benefits from a larger number of data constraints (Nx) that in UnetSurf rather than their subsurface location? I wonder if providing Nx time as many surface data to train UnetSurf would make it match the performance of Unet4layers.

- L220-223 : state how much of the training sets was "chaotic situations" versus nonchaotic ones. I wonder : if you trained a NN separately on just "chaotic situations", would you get improved performance? maybe match the performance obtained for nonchaotic ones.

- L294 : it seems to me that a confidence index would need to correctly account for several sources of model error (uncertain sinking velocity distributions, possible biases in POLGYR statistics, transport rates, etc) to avoid being misleading. Please discuss this in a bit more detail.

- in section 5 please describe the kind of real world observational experiment and data sets that would be needed to demonstrate / quantify the method's skill outside of a twin experiment configuration. I feel that many  would be concerned with using real world results of such a method if its skill can only be assessed within a model world.

technical corrections

L57 : "used to design"

L94 : space missing in "1000 m.Biological"

L96 : spell out what 3D+t means. also, consider saying 4D instead of 3D+t or 3D+T

L100 : "To avoid common particle between two experiments" is unclear and seems grammatically incorrect

L101 : rephrase as e.g. "to set up the patch centers 36km apart"

L104 : nx,ny would seem a better notation than deltax, deltay; no?

L109 : "for storage constrain" is vague and seems grammatically incorrect. 

Fig 3 caption : "superimposed" (one "s" only)