This paper describes a deep neural network to reconstruct subsurface thermohaline profiles from a combination of in-situ measurements and satellite data, focused on the Northwestern Pacific.

The technique by itself presents very relevant and innovative solutions to extract information from time series of satellite-based sub-images (centered on the target profile location) and combine them to correct available climatological data. However, the manuscript presents some misleading statements and some questionable claims that need to be corrected before publication. Additionally, several technical clarifications are also necessary, as detailed in the following.

Major remarks:

• In different sections, the authors cite “geostrophy” as a limiting factor in designing 3D reconstructions based on satellite altimetry. This is a conceptual imprecision, as geostrophy is an approximation used to derive horizontal currents from horizontal pressure gradients and has no influence on the sea level observations (which reflect all processes driven by full dynamics). Moreover, while it is commonly stated by the community that the sea surface elevation represents an integral measure of subsurface processes, this is actually true only in purely baroclinic flows and it is not true in general. Surface elevation is useful to infer the vertical distribution of hydrographic properties because the ocean is predominantly baroclinic and surface intensified (so that pressure differences are detectable at the surface). While this is discussed in section 4, it is not correctly introduced in the Abstract. Consequently, the second and third points listed in section 4 are redundant and should be merged removing any reference to non-geostrophy and eventually referring to non-baroclinic (the two terms should not be used interchangeably).
• the FiLM layer is not detailed to the level that would allow replicating the study. In fact, the neural network that modulates the scaling and shifting factors (conditioning network) is described nowhere (number of layers/neurons, etc.). It might also be worth adding a dedicated panel in figure 3.
• I understood the target of the TS-cast regression model is represented by the observed T/S profiles (2 channels, 128 layers). What is the “Error Variance” 3x128 tensor in figure 3? How is it computed and what are the three channels? Are the predicted/target variables normalized before entering the model? If yes, how (full profile standardization, depth-by-depth, min/max., etc.)? How does this eventually impact the loss functions defined in section 2.4?
• Following from previous point: in Section 2.4 the authors write: “TS-Cast network directly estimates temperature (ˆT) and salinity (ˆS) profiles, but also their associated depth-dependent uncertainty, represented by the logarithmic variance (log σ2)”. Are the authors speaking of the true variance or error variance? How is this practically computed? I suggest the terms “uncertainty” and “variance” are used much more carefully throughout the text as they are not synonyms. Again, please provide formulae to explain your operational definitions. In the present form, this section is very confusing.
• By the way, if the authors are relying on variance and not error variance, the Loss function used forces the model to learn more from the less dynamically active depth levels, which seems quite a sub-optimal choice, requiring further justification. In fact, this seems in contrast to the objective of the model, which would likely be to “resolve” the most relevant dynamics.
• The physics-informed loss function accounting for the equation of state has been first introduced in Sammartino et al., https://doi.org/10.1016/j.envsoft.2025.106660, and that paper should thus be cited in this section.
• How did the authors compute coherence?
• The authors' critique of what they name 'closed-loop' validation may be overly broad (both in the introduction and in the concluding remarks). Validation against independent data from the same platform (e.g., held-out Argo floats) is a standard and robust practice. While time-series validation at mooring sites provides valuable complementary information on temporal dynamics, it has its own spatial limitations. I suggest reframing this as a complementary strength of the present work rather than a limitation of prior approaches.

Minor points:

Line 1: “estimates of ocean surface states” -->“estimates of ocean surface state”

Line 3: “at sufficient space and time scale” --> “at sufficient space and time resolution”

Line 4: “While ADT represents” --> “While ADT contains”

Line 13 and following: “data-assimilated” --> “data-assimilating”

Line 18: remove “fundamental variables,”

Line 28: remove “primarily”

Line 39: change to “like multilinear regressions combined with optimal interpolation”

Line 44: change to “designed to capture sequential dependencies” (at least the first paper is not using LSTM over time, but over depth)

Line 46: add reference to Sammartino et al. (2025) - see above

Line 61 and following: considering the model domain, what is the advantage of transforming Lat-lon coordinates to 3 different variables?

L262: “explains”-->”might explain”

Line 284: “non-geostrophic”-->”non-baroclinic”