acceptodds
Withdrawn from ICLR 2027

CALM: Carbon-Aware Learning for Marine Chlorophyll Reconstruction in Three Dimensions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reconstructing three-dimensional ocean chlorophyll from sparse observations is essential for understanding marine ecosystems and carbon cycling. Direct tracer reconstruction overlooks that chlorophyll is a regulated pigment whose abundance depends on both phytoplankton carbon biomass and pigment allocation. We introduce CALM, a deep learning framework that reconstructs chlorophyll through coupled organic and inorganic carbon states. Conditioned on sparse chlorophyll, dense inorganic carbon, and environmental drivers, separate effective transport operators preserve each carbon pool’s global inventory, learned interconversion preserves their local sum, and surface and lower-boundary corrections account for external exchange. A constrained neural decoder uses a variable chlorophyll-to-carbon ratio to ensure that pigment carbon does not exceed available organic carbon. These properties follow from the architecture without conservation penalties. Experiments across three observation densities show improved reconstruction accuracy over the evaluated baselines, supported by carbon-budget audits and component ablations. At 10% sampling, CALM reduces NRMSE and RMSE by 11.6% and 19.5% over the strongest baseline.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Informed Deep Learning, 3D Chlorophyll Reconstruction, Biogeochemical Consistency

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_2ujaOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The overarching goal of the paper is to reconstruct 3D chlorophyll concentration in the world's oceans with constrains from DIC and organic carbon based on modeling data. The authors decompose this problem according to traser-conservation equation, i.e., temporal change is equal to transport plus source-sink term. Chlorophyll is associated with the source-sink term. The paper modeled the chlorophyll-to-carbon ratio to consider the physiology.

Critical strengths

The paper tried to model chlorophyll-to-carbon ratio.

Critical weaknesses

The paper is based on the assumption that carbon (DIC, POC, and DOC) have denser measurements than chlorophyll, which motivates the incorporation of the data to constrain 3D chlorophyll. However, in the ocean, there are a limited number of DIC, POC, and DOC (particularly these three together). Also, the conversion between DIC and organic carbon is controlled by many factors, e.g., heterotrophic respiration which is not directly associated with phytoplankton.

Questions for authors

I would recommend the authors to get more feedbacks from ocean biogeochemists.

Reviewer_v8ErOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CALM for reconstructing three-dimensional ocean chlorophyll from sparse chlorophyll observations, dense dissolved inorganic carbon, and environmental inputs. It combines a spatiotemporal backbone with globally inventory-preserving carbon transport, locally conservative organic–inorganic conversion, boundary corrections, and a bounded chlorophyll-to-carbon decoder. Experiments on monthly ocean simulations across three sampling densities report improvements over multiple baselines, supplemented by component ablations, carbon-budget audits, uncertainty estimates, and adaptation to an interpolated Argo-derived product.

Critical strengths

CALM distinguishes pigment reconstruction from carbon accounting through explicit transport, conversion, boundary, and decoding operations. The conservation and positivity properties are straightforward to inspect, and the paper carefully distinguishes global inventory preservation from local numerical conservation.

Critical weaknesses
  1. The experiments do not isolate the contribution of carbon awareness from the TIDE-UNet3D backbone and optimization choices: the factorized baseline differs architecturally, while the backbone ablation retains all carbon modules. Moreover, Table 2’s single-seed results show no CHL NRMSE improvement from conversion or DIC supervision, leaving the central mechanism’s predictive benefit unresolved.
  2. The Argo experiment interpolates each target field before selecting sparse inputs, allowing withheld original values to influence those inputs, and approximately 98% of the expanded targets are interpolated (Appendix E.2). Although explicitly acknowledged, this substantially limits the experiment as evidence of reconstruction on independent observations.
Questions for authors
  1. Are existing results available for a direct-output or unconstrained factorized decoder using the same TIDE-UNet3D backbone, inputs, and optimization as CALM?
  2. What are the learned OC magnitudes and chlorophyll-to-carbon ratio distributions, and are any corresponding simulation carbon fields available for validation? How do the authors assess whether the very small pigment-carbon fractions reflect meaningful pool estimates rather than compensating latent scales?
  3. Are any Argo results available where original observations are withheld before interpolation, with evaluation restricted to those untouched observations? If not, please clarify what evidence supports transfer beyond reconstruction of the interpolated product.
Reviewer_ETDpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Chlorophyll-a is a proxy for estimating phytoplankton biomass, tracing marine food web, and tracking global carbon sequestration. Satellite data only observes the ocean surface layer, and deep subsurface profiles are sparse, whereas the goal of this work is 3D reconstruction.

The works presents a deep learning framework that reconstructs chlorophyll by organic and inorganic carbon states using a 3D UNet.

The proposed model separates between global conservative transport, local conservative interconversion, and boundary exchange. It uses a constrained decoder and on a benchmark with simulated CHL and DIC labels, constraining carbon transport and interconversion structurally and decodes the pigment with a bounded ratio.

Critical strengths
  1. The work is well motivated. Mapping Chlorophyll-a is important for quantifying marine productivity, tracking ecosystem dynamics, and understanding ocean carbon sequestration.

  2. The ablation studies, removing components (such as diffusion, boundary source changes, DIC supervision, etc) while keeping all other factors equal are valuable.

  3. The constraints are built into the network rather than minimizing a penalty term.

Critical weaknesses

This work proposes to introduce CALM as a new deep learning framework for carbon-aware learning for marine chlorophyll.

It is missing key related work:

@article{niu2023incorporating, title={Incorporating marine particulate carbon into machine learning for accurate estimation of coastal chlorophyll-a}, author={Niu, Jie and Feng, Ziyang and He, Mingxia and Xie, Mengyu and Lv, Yanqun and Zhang, Juan and Sun, Liwei and Liu, Qi and Hu, Bill X}, journal={Marine Pollution Bulletin}, volume={192}, pages={115089}, year={2023}, publisher={Elsevier} }

This related work integrates marine particulate organic carbon (POC) and particulate inorganic carbon (PIC) to accurately map chlorophyll-a.

Limited novelty over previous work including Zheng et al, 2025.

Questions for authors

What is the relationship between this work and CALM acronym and Niu et al, 2023?

Reviewer_oUJ6OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents CALM, a carbon-aware framework for reconstructing three-dimensional ocean chlorophyll fields from sparse observations. Instead of directly predicting chlorophyll, the method models organic and dissolved inorganic carbon through separate transport operators, learned carbon interconversion, and boundary corrections. A constrained decoder then estimates chlorophyll using a variable pigment-to-carbon ratio. The authors evaluate CALM on a monthly ocean simulation dataset under three observation densities and compare it with several reconstruction baselines. The results show improved reconstruction accuracy, while additional experiments examine carbon conservation, model components, predictive uncertainty, and transfer to an Argo-derived chlorophyll product.

Critical strengths
  1. Reconstructing 3D ocean chlorophyll from sparse measurements is an important problem with practical scientific relevance. The paper addresses a real challenge in ocean monitoring, and the motivation for incorporating carbon-related information is reasonable.

  2. One positive aspect is that carbon conservation and interconversion are handled through the model structure rather than additional penalty terms. This makes the intended constraints easier to verify and avoids relying entirely on loss-weight tuning.

Critical weaknesses
  1. My main concern is that the proposed framework largely combines existing ideas, including neural transport, conservation constraints, and a factorized decoder. While their integration is reasonable, I am not fully convinced that the paper introduces a sufficiently new learning principle or methodological insight for ICLR.

  2. In Table 2, removing carbon conversion slightly improves chlorophyll NRMSE, while removing boundary corrections has very little impact. This makes it difficult to tell whether the reported accuracy gains actually come from the carbon-aware mechanisms or from the backbone and other architectural choices.

  3. The conservation properties are mathematically reasonable, but they mainly ensure consistency within the model's own predicted carbon states. Since organic carbon is not directly supervised, it remains unclear whether these states correspond to realistic ocean carbon distributions. In other words, satisfying a conservation identity is not the same as recovering the correct physical state.

  4. Most of the evaluation relies on a single MOM6–COBALTv2 simulation dataset. Although the paper includes an Argo-derived experiment, the target fields are interpolated before evaluation, which limits the evidence for generalization to independent real-world observations. I would like to see stronger validation under realistic observation conditions or across different ocean models.

  5. The practical benefit of some constraints is not demonstrated, For example, the unbounded decoder also achieves 100% biochemical consistency in the reported experiments. This raises the question of whether the explicit pigment-carbon bound provides a meaningful empirical advantage in the current setting. Similarly, some conservation-related components appear to contribute more to formal guarantees than to measurable reconstruction improvements.

Questions for authors
  1. Could the authors clarify which part of CALM contributes most to the improvement over UNet3D? In particular, would it be possible to compare the carbon-aware components against a matched TIDE-UNet3D model without the proposed physical constraints?

  2. Table 2 shows that removing carbon conversion slightly improves chlorophyll NRMSE, while removing boundary corrections has almost no effect. How should these results be interpreted in relation to the claimed benefits of carbon-aware reconstruction?

  3. Since organic carbon is not directly supervised, how can the authors determine whether the reconstructed organic carbon fields are physically meaningful rather than simply latent representations that satisfy the imposed constraints?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.