acceptodds
Desk rejected at ICLR 2027

PhysiOcean: A Spatio-Temporal Operator with Physics-Guided Learning for Coastal Dissolved Oxygen Forecasting

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Forecasting dissolved oxygen (DO) across monitoring locations and depth layers is challenging because coastal DO dynamics involve multiscale environmental forcing, strong spatial heterogeneity, and sparse observations. Existing data-driven approaches typically model individual stations independently or rely on fixed spatial decompositions, limiting their ability to jointly represent temporal dynamics and continuous spatial variation. We introduce , a spatio-temporal operator that maps historical environmental forcing and spatial coordinates to DO concentrations. PhysiOcean combines a multi-scale causal temporal encoder, a learnable Fourier spatial decoder, and a multi-head nonlinear fusion mechanism to learn regional DO fields end-to-end. EOF-guided initialization and physics-guided regularization further incorporate dominant spatial structure and domain knowledge during training. We analyze the expressivity of the proposed fusion mechanism and the complexity of an idealized physics-constrained hypothesis class. Experiments on a 32-year monitoring record from a large, environmentally heterogeneous estuary containing 40 monitoring stations and 800 station-depth locations show that PhysiOcean outperforms sequence, decomposition-based, and operator-learning baselines. Component analyses demonstrate that the temporal encoder, continuous spatial representation, fusion mechanism, EOF initialization, and physics-guided losses provide complementary improvements. PhysiOcean also maintains stronger performance at monitoring locations and depth layers excluded from training, demonstrating within-estuary generalization to unobserved spatial coordinates. Our code is available at https://anonymous.4open.science/r/physiocean/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Informed Learning, Neural Operators, Dissolved Oxygen Forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hLXgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The overarching goal of this paper is to forecast dissolved oxygen of water column from environmental variables. The authors developed this algorithm using observations from a local region (Chesapeake Bay). The major technical contribution of this work is the incorporation of physical regularization in the final loss function.

Critical strengths

The major technical contribution of this work is the incorporation of physical regularization in the final loss function.

Critical weaknesses

My major comments are listed below. First, the first physical regularization (i.e., the oxygen is forced undersatulation) will be violated in the surface ocean in the summer due to biology. This component should be large when the biological oxygen is large. Second, the second physical regularization will be the same. The source-sink term should be important in some months and depths. Third, the amount of training data compared to the number of parameters (~650K) is very small. I am not sure whether overfitting is problem. Fourth, the forcing variables are not clear. I can only find river discharge,nutrient loading,air temperature,wind,and solarradiation. Apparently, other important variables for dissolved oxygen are not included here. The water temperature, rather than air temperature, will dominate dissolved oxygen. Salinity also influences dissolved oxygen. Another variable will be the biological parameter (e.g., phytoplankton).

Questions for authors

No

1 comment

We sincerely thank the reviewer for the positive evaluation and constructive feedback. We appreciate the recognition of the value of physics-guided regularization for coastal DO forecasting. We address the four concerns below.

1. Oxygen saturation constraint and biological supersaturation. We agree that biological activity, particularly photosynthesis, can cause oxygen supersaturation in surface waters during summer. Importantly, our approach employs a soft physical constraint rather than strictly enforcing DO concentrations below saturation. Specifically, when temperature and salinity data are available, we calculate the oxygen saturation concentration using �0��0�; otherwise, we adopt a relaxed fixed upper bound of 14.0 mg/L. Since this constraint is incorporated as a soft regularizer, the model can still predict DO concentrations exceeding the saturation threshold when supported by observations. Its purpose is to discourage physically implausible predictions rather than assume universally undersaturated conditions. We will clarify this implementation and discuss its limitations under intense biological activity.

2. Spatial smoothness and biogeochemical source–sink processes. We acknowledge that localized biological production and consumption can generate substantial spatial DO gradients. However, our smoothness penalty does not enforce spatially uniform DO fields. Instead, it serves as a weak regularizer that suppresses unsupported spatial oscillations while allowing the data-fitting objective to preserve meaningful spatial heterogeneity. Our hyperparameter sensitivity analysis further shows that excessive smoothing degrades prediction accuracy, whereas moderate regularization improves spatial generalization. Thus, our approach balances observational fidelity and spatial regularity rather than assuming negligible source–sink effects. We will further clarify the physical interpretation and applicability of this constraint.

3. Training data size and overfitting. We clarify that our training data cover 288 months across 800 spatial locations, yielding approximately 230,400 spatiotemporal supervised observations, rather than only 288 monthly samples. Although these observations are not statistically independent, PhysiOcean learns shared spatiotemporal representations through its Branch–Trunk architecture instead of fitting separate models at individual locations. We also employ weight decay, early stopping, and physics-guided regularization to mitigate overfitting. More importantly, PhysiOcean achieves R² = 0.881 on the independent 2009–2016 test period and demonstrates generalization to stations and depth layers excluded from training. These results provide empirical evidence of generalization beyond the training observations. We will additionally provide training and validation loss curves to further assess overfitting.

4. Selection of environmental forcing variables. We agree that water temperature, salinity, and biological variables such as phytoplankton are important drivers of DO dynamics. However, our study focuses on forecasting continuous DO fields from regionally aggregated external environmental forcings, rather than relying on local water-quality observations at individual monitoring locations. Our 39 forcing variables include air temperature, river discharge, nutrient loading, wind, precipitation, and solar radiation, which indirectly characterize thermal, transport, and biogeochemical processes. Importantly, temperature and salinity data, when available, are used to calculate oxygen saturation for physical regularization and should not be confused with the environmental forcing inputs to the Branch Network. We acknowledge that external forcings cannot fully replace local water temperature, salinity, and biological observations. We will provide the complete list of forcing variables, clarify the distinction between forecasting inputs and auxiliary variables used for physical constraints, and discuss incorporating local environmental and biological observations in future work.

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

PhysiOcean adapts a DeepONet branch–trunk framework for coastal dissolved oxygen (DO) prediction from environmental forcing histories and spatial coordinates. It combines a multiscale causal temporal encoder, a Fourier-feature spatial trunk, headwise fusion with a nonlinear correction, EOF-guided initialization, and four physics-guided soft regularizers. Experiments on Chesapeake Bay evaluate temporal prediction, station/depth holdouts, and component ablations. The spatial claims are appropriately limited to the study estuary.

Critical strengths
  1. PhysiOcean adapts a DeepONet branch–trunk framework for coastal dissolved oxygen (DO) prediction from environmental forcing histories and spatial coordinates. It combines a multiscale causal temporal encoder, a Fourier-feature spatial trunk, headwise fusion with a nonlinear correction, EOF-guided initialization, and four physics-guided soft regularizers. Experiments on Chesapeake Bay evaluate temporal prediction, station/depth holdouts, and component ablations. The spatial claims are appropriately limited to the study estuary.
  2. The experiments use a 32-year record, a chronological test split, five random seeds for the main comparisons, station/depth holdouts, and component ablations. The reported improvements over EOF+NN and several forecasting baselines support the practical usefulness of the combined framework within Chesapeake Bay. The physical-consistency diagnostics show improved adherence to the specified priors, while evaluation at withheld coordinates tests a capability beyond prediction at previously observed locations. These results provide useful application evidence, even though attribution of the gains remains incomplete.
Critical weaknesses
  1. PhysiOcean builds on the established branch–trunk factorization and integrates familiar temporal convolutions, attention, Fourier spatial features, EOF-based initialization, and soft physical regularizers. These design choices are coherent and may be effective, but their combination alone does not clearly establish a distinct operator-learning methodology. More specifically, the weighted sum of head-wise inner products remains a reweighted bilinear pairing because the gate weights are input-independent. Given existing work on nonlinear operator decoders and POD/EOF-based operator representations, the paper should more clearly establish what is conceptually new about its fusion and physics-guided formulation, beyond a task-specific integration of known techniques.

  2. The recent baselines are drawn mainly from time-series forecasting rather than neural operator learning, and several additionally use target-location DO histories. They are useful practical comparisons but do not directly assess the operator architecture. The method named DeepONet in Table 1 retains the proposed architecture and removes only EOF pre-training and physics losses. Comparisons with representative operator methods—such as a separately tuned standard DeepONet, POD-DeepONet, or a suitable nonlinear-decoder variant—under matched inputs and spatial hold-out protocols would be needed to substantiate the claimed advance over prior operator-learning methods.

  3. Although the paper provides quantitative evaluations and visualizes temporal attention and learned spatial bases, it lacks direct spatiotemporal comparisons between predicted and observed DO fields. For a spatially heterogeneous environmental forecasting task, visualizations of seasonal DO distributions, vertical profiles, and spatial error patterns would help assess whether PhysiOcean captures meaningful physical structures and their temporal evolution. The current presentation makes it difficult to understand where and under what conditions the model performs well or fails, beyond aggregate metrics.

Questions for authors
  1. Could the authors further clarify the methodological novelty of the proposed multi-head fusion and physics-guided learning compared with existing nonlinear DeepONet extensions and physics-informed operator learning approaches?
  2. Could the authors include representative neural operator baselines, such as POD-DeepONet, under comparable experimental settings to better substantiate the proposed method's advantages?
  3. Could the authors provide spatiotemporal visualizations comparing predicted and observed DO fields, including seasonal patterns, vertical structures, and spatial errors, to better demonstrate the model's ability to capture meaningful physical dynamics?
3 comments

We sincerely thank the reviewer for the thoughtful evaluation and constructive feedback. We appreciate the recognition of the practical value of PhysiOcean and its evaluation on long-term coastal DO observations. We address the concerns below and will incorporate the suggested clarifications and additional experiments into the revised manuscript.

1. Methodological novelty (W1 & Q1). We acknowledge that temporal convolutions, Fourier features, and physics-based regularization have been explored in prior studies. However, PhysiOcean's primary contribution is not a collection of novel components, but a unified physics-guided spatiotemporal operator for continuous DO field forecasting. Specifically, we extend classical EOF decomposition from fixed spatial modes to learnable continuous spatial representations, using EOF modes for data-driven initialization while allowing the spatial bases to adapt through end-to-end optimization. Combined with nonlinear multi-head fusion and four domain-specific physical constraints, this framework jointly learns environmental forcing responses and spatial structures under limited observations. We agree that fixed-gate aggregation remains bilinear, but the complete fusion mechanism includes a nonlinear correction beyond this formulation. We will clarify this methodological contribution and its distinctions from existing nonlinear DeepONet and physics-informed operator approaches.

2. Neural operator baselines (W2 & Q2). Our experiments evaluate both practical forecasting performance and component effectiveness. Although some recent forecasting baselines additionally use target-location DO histories, they provide meaningful practical comparisons. The DeepONet baseline in Table 1 retains the proposed architecture while removing EOF initialization and physics losses, specifically to assess their combined contribution rather than represent an independently implemented standard DeepONet. We agree that additional operator-learning comparisons would strengthen the evaluation. We will include independently tuned standard DeepONet and POD-DeepONet baselines under matched environmental inputs, spatial coordinates, data splits, and spatial hold-out protocols to ensure fair comparisons.

3. Spatiotemporal DO field visualization (W3 & Q3). We agree that aggregate metrics and learned-basis visualizations alone cannot fully demonstrate the reconstruction of meaningful DO structures. We will add direct comparisons between predicted and observed DO fields, including seasonal spatial distributions, representative vertical DO profiles, and spatial prediction error maps. These visualizations will illustrate how PhysiOcean captures seasonal variability, vertical stratification, and spatial heterogeneity, while also revealing prediction errors and limitations under challenging environmental conditions.

Thank you for the detailed response. I appreciate the authors' clarification of the methodological contributions and their willingness to address the concerns raised in the review.

The clarification of the framework is helpful. However, I still have some questions. It is not yet clear how the framework is methodologically distinct from existing nonlinear operator-learning approaches. I welcome the proposed additions of Neural Operator baselines, as well as spatiotemporal DO visualizations. These additions would directly address my main concerns.

However, since the additional experiments and visualizations have not yet been presented, it is difficult to assess whether they sufficiently strengthen the paper's contributions. I would be interested in seeing these results in the future version of manuscript. For now, I maintain my original assessment of weak reject.

We thank the reviewer for the follow-up feedback. Despite the limited rebuttal period, we conducted additional experiments with two random seeds to further validate our design.

1. Learnable spatial representations. Under the same 10-dimensional setting, the learnable EOF-initialized Trunk achieves higher R² than the frozen Trunk (0.8627 ± 0.0080 vs. 0.8569 ± 0.0048), supporting the benefit of adaptive spatial bases.

2. EOF–physics interaction. Our four-configuration ablation shows that combining EOF initialization and physics losses achieves the highest R² (0.8727 ± 0.0086), supporting their potential positive interaction rather than merely independent contributions.

We acknowledge that these results do not replace comparisons with existing neural operators. We will add standard DeepONet and POD-DeepONet baselines, along with spatiotemporal DO visualizations, in the revised manuscript to further substantiate our contributions.

Reviewer_DY5XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper designs PhysiOcean -- a method for forecasting monthly dissolved oxygen (DO) across Chesapeake Bay (40 stations × 20 depths) from bay-wide drivers such as river flow, nutrient loads, and weather. PhysiOcean extends the classic EOF approach, where DO is computed as the product of fixed spatial maps and time-varying weights, by learning both the spatial maps and the time-varying weights within a DeepONet. A convolutional, attention-based branch network produces the time-varying weights from 24 months of drivers. A Fourier-feature trunk network turns coordinates into continuous spatial maps that can be queried at unmeasured locations. This trunk network is initialized by pre-training on the classical EOF-based spatial maps. The authors also add a nonlinear multi-head fusion step and four soft physics penalties as constraints. PhysiOcean outperforms EOF-based, recurrent, recent forecasting, and DeepONet. The authors present ablations to examine each component.

Critical strengths
  1. The paper addresses forecasting coastal dissolved oxygen across a whole estuary, using a 32-year monitoring record. This is an environmental problem that the ML community rarely studies.
  2. The branch/trunk design maps onto the classic EOF method -- fixed spatial maps multiplied by time-varying weights. Using the most recent month’s features as the attention query over the 24-month history is a sensible way to produce the temporal weights. Because the spatial maps are a continuous function of coordinates, the model can be queried at stations and depths not seen in training within the training estuary.
Critical weaknesses
  1. The trunk and EOF initialization may be limited in novelty than the paper suggests. Both are listed as contributions, but close precedents exist: FEDONet [1] places fixed, non-learnable Fourier features inside a DeepONet trunk. POD-DeepONet [2] uses POD modes of the training data (POD is equivalent to EOF) as the trunk basis. Lee & Shin [3] propose a two-step DeepONet training scheme that trains the trunk first and then the branch.

  2. Some ablation results are internally inconsistent. “No EOF pre-training” is essentially the DeepONet baseline plus the physics losses, but it scores lower (R² 0.841, Table 2 vs 0.861, Table 1). This indicates the physics losses hurt without EOF pre-training. The paper does not discuss this. Table 2 also reports no standard deviations. Multi-seed ablations with comparable tuning are needed before the component contributions can be trusted.

[1] Sojitra, Arth, Mrigank Dhingra, and Omer San. "FEDONet: Fourier-embedded DeepONet for spectrally accurate operator learning." Journal of Computational Physics (2026): 114931. [2] Lu, Lu, et al. "A comprehensive and fair comparison of two neural operators (with practical extensions) based on fair data." Computer Methods in Applied Mechanics and Engineering 393 (2022): 114778. [3] Lee, Sanghyun, and Yeonjong Shin. "On the training and generalization of deep operator networks." SIAM Journal on Scientific Computing 46.4 (2024): C273-C296.

Questions for authors
  1. The abstract names components such as the “learnable Fourier spatial decoder” and “EOF-guided initialization” without explaining them. The key intuition only becomes clear later on in the paper, that PhysiOcean is a learned version of the classic EOF approach. Would the authors consider stating this connection in the abstract and introduction? It would make the architecture much easier to follow.

  2. What motivates using convolutions in the branch network, and why are kernel sizes 2, 4, and 8 with dilations 1 and 2 appropriate for monthly DO data?

  3. Table 2 removes EOF pre-training entirely, but no experiment keeps the EOF-fitted trunk fixed. That comparison is necessary for the claim that learned maps beat fixed EOF maps. Could the authors report results on the standard split and on the held-out station and depth settings for: (i) a frozen EOF-fitted trunk with p = 10; (ii) the first 10 bases frozen, with the remaining 54 trainable.

  4. The comparison model in Figure 2 (“PhysiOcean-no-physics”) also lacks EOF pre-training (Section 4.2). But Section 4.2 attributes the hold-out gains to the physics losses only. Could the authors provide separate hold-out curves for “no physics loss” and “no EOF pre-training”?

3 comments

1. Novelty of EOF initialization and spatial learning. We acknowledge prior work on Fourier features, POD/EOF representations, and two-stage training. However, PhysiOcean's contribution lies in extending classical EOF decomposition into an end-to-end learnable spatiotemporal operator. Unlike FEDONet's fixed Fourier features and POD-DeepONet's POD-based representation, our framework employs learnable multi-scale Fourier features and uses EOF modes for initialization rather than fixing the spatial basis. Moreover, nonlinear multi-head fusion extends beyond the bilinear formulation of standard DeepONet. We will clarify these distinctions.

2. Ablation consistency and component interaction. We emphasize that the results are not inherently contradictory, as they correspond to different initialization conditions. Table 2 directly demonstrates the positive contribution of physics losses: with EOF initialization retained, adding physics regularization improves R² from 0.8674 to 0.8814. The lower performance without EOF initialization suggests a potential positive interaction between these components. EOF pre-training provides a structured spatial starting point, allowing physics constraints to guide subsequent optimization more effectively. Without it, physical regularization may interfere with immature spatial representations. We will clarify this interaction and report multi-seed ablations with standard deviations under consistent settings.

3. Connection to classical EOF. We agree that this connection should be introduced earlier. PhysiOcean extends classical EOF into a learnable operator: the Branch Network predicts temporal coefficients, the Trunk Network learns continuous spatial basis functions, and nonlinear fusion models their interactions. We will explicitly state this intuition in the abstract and introduction.

4. Convolutional architecture. Multi-scale causal convolutions capture environmental forcing effects across different temporal scales. Kernel sizes 2, 4, and 8 with dilations 1 and 2 yield theoretical receptive fields of 8, 22, and 50 months, respectively. The largest field is effectively limited to the 24-month input window. This design captures short-term, seasonal, and longer-range dependencies efficiently. We will clarify the architectural motivation and parameter selection.

5. Fixed versus learnable EOF bases. Fixed EOF modes cannot adapt to the forecasting objective, whereas PhysiOcean uses EOF only for initialization and subsequently refines spatial bases through end-to-end learning. Existing results support this design: PhysiOcean achieves R² = 0.881 versus 0.848 for EOF+NN, while removing EOF initialization reduces R² to 0.8412. These findings support the overall framework, although a controlled frozen-basis comparison would further isolate the contribution of spatial learning. We will evaluate a frozen EOF trunk (p=10) and a partially frozen configuration (10 fixed + 54 trainable bases) under standard and spatial hold-out settings.

6. Separating physics and EOF contributions. We acknowledge that Figure 2's “PhysiOcean-no-physics” removes both physics losses and EOF pre-training, so it evaluates their combined effect. As discussed above, EOF initialization and physics constraints may be complementary: the former provides meaningful spatial structure, while the latter encourages physically consistent predictions. We will correct the labeling and provide separate held-out station and depth curves for each ablation, distinguishing their individual contributions and potential interaction.

Thank you for the detailed response. I appreciate the commitments to state the EOF connection early, explain the branch design, and correct the Figure 2 labeling.

My two main concerns remain open, because the response relies on promised rather than reported results:

  • Ablation consistency: The proposed interaction between EOF initialization and the physics losses is plausible, but it is untested. The gain cited as support (0.8674 → 0.8814) is itself a single-run Table 2 difference without a standard deviation. Could you share the multi-seed ablations if possible?
  • Fixed vs. learned bases: The frozen-trunk experiments would directly test the paper’s central claim. Without EOF initialization, PhysiOcean scores R² = 0.841 (Table 2), which is below the fixed-EOF baseline EOF+NN at 0.848 (Table 1). I recognize these come from different evaluation protocols, but they make the controlled comparison important.

I will keep my score for now and am happy to revisit it if these results are provided.

We sincerely thank the reviewer for the follow-up comments and appreciate their emphasis on experimental validation. Given the limited rebuttal period, we prioritized additional ablation and frozen-trunk experiments using two additional random seeds, reporting the mean and standard deviation. These new results provide further empirical support for our previous explanations.

1. Multi-seed ablations and the interaction between EOF initialization and physics losses.

We conducted additional experiments evaluating all four combinations of EOF initialization and physics-guided regularization:

EOFPhysicsR² (Mean ± Std)RMSE (mg/L, Mean ± Std)
OffOff0.8621 ± 0.00621.0899 ± 0.0245
OnOff0.8637 ± 0.00341.0835 ± 0.0135
OffOn0.8575 ± 0.01511.1069 ± 0.0581
OnOn0.8727 ± 0.00861.0470 ± 0.0355

These results provide further evidence of a positive interaction between EOF initialization and physics losses. With EOF initialization, adding physics regularization improves R² from 0.8637 to 0.8727 (+0.0090), whereas without EOF initialization, physics losses do not yield the same improvement.

This suggests that EOF initialization provides a structured spatial starting point, enabling physics-guided regularization to guide subsequent optimization more effectively. The full PhysiOcean achieves the best average performance across all four configurations, supporting the effectiveness of combining these two training strategies.

2. Fixed versus learnable spatial bases.

We additionally conducted frozen-trunk experiments to directly assess the contribution of learnable spatial basis functions.

ConfigurationR² (Mean ± Std)RMSE (mg/L, Mean ± Std)
Frozen 64-dimensional trunk0.8231 ± 0.00111.2346 ± 0.0038
Frozen 10-dimensional trunk0.8569 ± 0.00481.1103 ± 0.0186
Learnable 10-dimensional trunk0.8627 ± 0.00801.0877 ± 0.0317

Under the same 10-dimensional basis setting, the learnable trunk achieves higher average R² than the frozen trunk (0.8627 vs. 0.8569), providing direct evidence that end-to-end adaptation of spatial basis functions can improve forecasting performance.

Furthermore, freezing the full 64-dimensional trunk leads to substantially lower performance (R² = 0.8231), further highlighting the importance of spatial representation adaptability.

These findings provide more direct empirical support for PhysiOcean's design of using EOF-guided initialization followed by end-to-end optimization, beyond the indirect comparison with the EOF+NN baseline.

We acknowledge that the additional experiments currently involve a limited number of random seeds and do not yet establish statistical significance. We will include more comprehensive multi-seed evaluations, the remaining frozen-basis configurations, and spatial hold-out comparisons in the revised manuscript, reporting means and standard deviations to further verify the robustness of these findings.

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

Related papers

Discussion (0)

Sign in to comment.