Physics-Encoded Inverse Modeling for Arctic Snow Depth Estimation

Akila Sampath; Jianwu Wang; Vandana P. Janeja

REVIEW 4 major objections 3 minor 13 references

Physics-Encoded Inverse Modeling for Arctic Snow Depth Estimation

T0 review · 4 major / 3 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read PhysE-Inv embeds a hydrostatic-balance proxy inside an LSTM encoder-decoder and reconstructs Arctic snow-depth proxies with 24.4% lower MSE than four capacity-matched baselines.

desk verdict The method is a competent re-combination of known pieces, but the evaluation targets a self-generated proxy built from inputs the model never sees, so the reported gains say nothing about real snow depth. read the letter →

arxiv 2601.17074 v5 pith:CDUWTVLZ submitted 2026-01-23 cs.LG cs.AI

Akila Sampath , Vandana P. Janeja , Jianwu Wang This is my paper

classification cs.LGcs.AI

keywords Arcticsnowdepthseaiceinversemodelingphysics-encodedneuralnetworkshydrostaticbalancecontrastivelearningLSTMencoder-decoderdata-scarceestimation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that Arctic snow depth, though absent from reanalysis products, can be recovered by framing the task as a surjective inverse problem and encoding the hydrostatic-balance equation directly into the learning pipeline. Since the true quantity is unobserved, the authors build an analytical proxy target from sea-ice concentration, snow albedo, and constant densities, and train an LSTM encoder-decoder with attention to invert noisy, sparse input time series toward that proxy. The framework's inverse head estimates three time-varying parameters per step, and a contrastive loss on noise-augmented latent states keeps the learned representations physically consistent. Under this proxy evaluation, PhysE-Inv reports an average MSE reduction of 24.4% over capacity-matched baselines and 17.3% over the strongest baseline. A sympathetic reader would take the contribution as evidence that simple, interpretable physics encoding, not a larger architecture, drives the gain.

What carries the argument

The load-bearing object is the hydrostatic-balance proxy, a three-term simplification of ice-thickness equilibrium: h_i ~ (rho_w * C + alpha * rho_s) / (rho_w - rho_i). It does double duty: it generates the training target, since direct snow-depth observations are not used, and it justifies the surjective inverse head that estimates per-step alpha, beta, and gamma. The surrogate reconstruction then combines the network's direct prediction and its sequence mean, with a contrastive loss keeping the latent representation stable under Gaussian input augmentation. The surjectivity assumption is what converts an ill-posed, non-unique inversion into a tractable many-to-one mapping.

What would settle it

Compare PhysE-Inv's predicted snow-depth proxy against independent snow-depth observations over the central Arctic, for example airborne or satellite altimetry retrievals not used in the reanalysis data. If the model tracks the proxy closely but diverges systematically from those measurements, the central claim is confined to the proxy. The paper provides no such comparison, so this check remains open.

Watch

Extended reading notes

Core claim

The central claim is that the inverse map from sparse, noisy observations to the physical target is many-to-one, and that modeling it as a surjective mapping with bounded, time-varying parameters produces better and more physically consistent reconstructions than bijective or unconstrained mappings. Concretely, the inverse head outputs (alpha, beta, gamma) transformed to ranges [-1,1], (0,infinity), and [-10,10], and the physics-encoded reconstruction combines the direct prediction and its sequence mean: h_est = alpha * h_pred + beta * mean(h_pred) + gamma. The reported results are MSE 0.3568 and RMSE 0.5973 with parameter estimation, versus the best baseline at 0.4315/0.6569, and the model'

Load-bearing premise

The target used for training and scoring is not measured snow depth but a simplified hydrostatic-balance proxy in which sea-ice concentration and albedo stand in for freeboard and snow depth; if that proxy is unfaithful, the reported error reductions do not transfer to real Arctic snow depth.

Editorial extensions

If this is right

Reported 24.4% average MSE reduction over four capacity-matched baselines, with 17.3% improvement over the strongest baseline in the parameter-estimation setting.
The framework reconstructs the seasonal cycle of the snow-depth proxy and captures abrupt transitions such as compaction or melt events with only minor phase lag or smoothing.
Contrastive learning yields lower MSE across training-data fractions down to 50%, indicating resilience to data sparsity.
Because the novelty is in the constrained inversion rather than the network topology, the same surjective inverse head can be attached to other sequence models and transferred to other data-scarce Earth-science inverse problems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The decisive untested step is external validation: scoring against the hydrostatic-balance proxy does not yet show skill at measured snow depth. A natural extension is to compare PhysE-Inv outputs against independent airborne or satellite snow-depth retrievals over the same region and period.
Because the reconstruction target is self-generated, one could train the same architecture on a different proxy, such as sea-ice thickness or freeboard, and check whether the alpha, beta, gamma parameters retain the same ranges, a quick test of whether the physics encoding is doing structural work or merely fitting.
Visualizing the learned parameters as a time series would turn the inverse head into an interpretable physical-state diagnostic, something the paper does not currently show.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things you should know. First, the paper is honest enough in its abstract: it says 'snow depth proxies,' but the title and conclusion claim snow depth estimation. Second, the evaluation target is built from variables (sea ice concentration, albedo) that are not in the model input (only snow density), and it's never checked against any real snow depth observations. So the numbers are uninterpretable as evidence about the actual quantity of interest.

What's genuinely there: the framework is a clean assembly of an LSTM encoder-decoder with attention, a contrastive loss, and a small parameter head that outputs bounded values. The authors are clear that the novelty is not the architecture but the 'constrained inversion' framing. They also make a fair point that bijective inverse mappings are too rigid for time-varying systems. The writing is mostly transparent about what was done, and they include an ablation on data sparsity.

The problems are load-bearing. Eq. 5 is not derived from Eq. 4; it replaces freeboard with a product of sea ice concentration and albedo with no units or physical argument that survives contact with the hydrostatic equation. The conceptual justification in §4.1.1 is hand-waving — albedo saturates after a few centimeters of snow, which is exactly why it is a poor proxy for depth. More damaging: Table 4 says the input is only the snow density time series, while the proxy target depends on SIC and albedo. Unless those are deterministic functions of snow density — which is not shown and is physically false — the target is not identifiable from the input. The model can only fit the conditional mean of the proxy given ρs, and the reported MSE reduction says nothing about how well it tracks the actual drivers of snow depth.

The reconstruction loss (Eq. 11) is also self-referential: it compares the direct prediction to an affine transform of that same prediction with learned α, β, γ. That constrains nothing independent; it's smoothness regularization wearing a physics costume. No error bars, no external validation, and the code link is not actually available in the version I read.

Who is this for? Someone working on physics-guided ML might read it as a cautionary example of proxy validation, but it is not a reliable source for Arctic snow depth estimates. My advice: desk reject or send back with a clear request for validation against independent snow depth (e.g., Operation IceBridge, CryoSat-2). The framework might be salvageable, but the central claim is unsupported as written.

Referee Report

4 major / 3 minor

Summary. The paper proposes PhysE-Inv, an LSTM encoder-decoder with multi-head attention and contrastive learning, for estimating Arctic snow depth from ERA5 reanalysis data. Because direct snow-depth observations are unavailable, the authors define a target 'snow depth proxy' starting from a hydrostatic-balance equation, then train and evaluate the model against this self-generated proxy. They report that PhysE-Inv outperforms LSTM, BiLSTM, Neural ODE, and ResNet baselines, with an average MSE reduction of 24.4% under the proxy evaluation framework. The central claim is that the method accurately reconstructs the seasonal evolution of snow depth and offers a general physics-encoded inverse-modeling approach for data-scarce geophysical problems.

Significance. If the target proxy were a faithful surrogate for real snow depth and the physics constraints were genuinely anchored to the hydrostatic balance, the framework would be a useful contribution to Arctic snow-depth estimation. The paper also brings a reasonable combination of sequential modeling, contrastive learning, and inverse parameter estimation. However, the current manuscript does not validate the proxy against any independent snow-depth observations, and several load-bearing steps in the derivation and loss construction are internally problematic. As it stands, the reported performance gains are against a self-generated target and cannot be interpreted as evidence about actual Arctic snow depth. The paper would need substantial revision—and, in particular, external validation and a corrected physical formulation—before its central claims could be accepted.

major comments (4)

[§4.1, Eq. (5)] Equation (5) does not follow from Eq. (4). Eq. (4) gives a length hi in terms of hs, fb, and densities. Eq. (5) replaces hs with α and fb with C, producing ρwC + αρs/(ρw−ρi). The first term has units of density and the second is dimensionless, so the right-hand side is not a length; the expression is dimensionally inconsistent. The conceptual justification in §4.1.1 argues that SIC and albedo are empirical proxies, but it does not explain how they enter the hydrostatic balance equation in this algebraic form, nor how the result can be a physically meaningful target for snow-depth prediction. Since this proxy defines the ground truth for all experiments, the validity of the central claim rests on this unvalidated and dimensionally questionable construction.
[§4.2, Eqs. (10)–(11)] The reconstruction proxy in Eq. (10) is an affine transformation of the model's own direct prediction, ĥest = α·ĥpred + β·ĥpred + γ (with the second term intended as a mean over the sequence). The loss L_PE-pred in Eq. (11) then measures the squared difference between ĥpred and this affine transform. This is a self-consistency regularizer; it does not enforce agreement with the hydrostatic balance equation or any independent physical relation. The learned α, β, γ are not identifiable as physical parameters, and the term 'physics encoding' is therefore misleading. The model could satisfy L_PE-pred by learning a trivial affine map, regardless of any physical law.
[§3.2 and Table 4] The paper states that the input is X = ρs (snow density time series), and Table 4 confirms the input feature is the snow density field only. The target proxy in Eq. (5), however, depends on sea-ice concentration C and snow albedo α in addition to ρs. No relationship is shown that would allow C and α to be determined from ρs. Consequently, the target is not identifiable from the input; the model can at best estimate the conditional mean of the proxy given ρs. This undermines the inverse-modeling formulation, which claims to estimate hidden physical parameters from observations that are not actually provided at the input.
[§6 and §4.1] All experimental results are evaluated against the self-generated proxy target; no external validation against independent snow-depth observations is presented. The conclusion states that PhysE-Inv 'accurately reconstructed the seasonal evolution of snow depth' and outperforms baselines, but this is only true for the proxy defined in Eq. (5). The paper itself lists 'validation against independent remote-sensing and in-situ sources' as future work. Without such validation, the reported 24.4% MSE reduction is uninterpretable as evidence about real Arctic snow depth. This is a load-bearing gap: the manuscript's title and conclusion frame the contribution as snow-depth estimation, not merely proxy fitting.

minor comments (3)

[Eq. (10)] The notation in Eq. (10) is ambiguous: both terms on the right are written with the same symbol ĥpred, although the text indicates the second is the mean of intermediate predictions. Please use an overline or a distinct symbol, e.g., ĥpred, to avoid confusion.
[Table 4] The test shape is listed as [2183, 10, 1], but the reader's report refers to train/test sizes as 8757/21. Verify the transcription in Table 4 and in the text; the actual test size appears to be 2183.
[Abstract and Table 1] The abstract quotes a 24.4% average MSE reduction, but the table reports per-model reductions that vary widely (e.g., vs. LSTM ~21.5%, vs. BiLSTM ~31.1%). State explicitly how the average is computed (which baselines and whether it is mean of relative reductions), so the number is reproducible.

Circularity Check

2 steps flagged · score 8.0 of 10

The 'physics-encoded' constraint is a self-consistency affine map on the model's own output, and the only evaluation target is a proxy generated from the same observational fields; the snow-depth claims therefore reduce to fitting a self-generated label.

self definitional [§4.2, Eq. 10–11]
"ˆhest s,t =α t · ˆh pred s,t +β t · ˆhpred s,t +γ t (10) where ˆhpred s,t is the direct prediction, and ˆh pred s,t is the mean of the intermediate predictions over the sequence lengthN=n+ 1. The corresponding loss,L PE-pred, ensures the final output respects the physical constraints: LPE-pred = 1 n+ 1 nX t=0 (ˆhpred s,t − ˆhest s,t)2 (11)"

The 'physics-constrained estimate' h_est is defined as an affine function of the model's own direct prediction and its mean. Minimizing L_PE-pred is trivially satisfiable by setting α=1, β=0, γ=0 (or α+β=1, γ=0), which makes h_est = h_pred and gives zero loss for any prediction. No hydrostatic-balance equation, Eq. 4, or proxy Eq. 5 enters this constraint. Thus the claimed 'physics encoding' is a self-consistency identity, not a physical law; any reported physical-consistency gain is available by construction.
fitted input called prediction [§3.1, §4.1 (Eq. 5), §6, A.6]
"To prepare the dataset for the inversion process, we applied spatial averaging and normalization to key parameters, including snow albedo, snow density, and sea ice concentration. These variables were subsequently processed through a proxy model based on the hydrostatic balance equation to estimate the target hidden characteristics of the sea ice. ... Input Feature Snow density field (time series); Target Normalized snow depth proxy ... Future work will incorporate Bayesian uncertainty quantification to validate learned parameters against independent remote-sensing and in-situ sources."

The target used for training and evaluation is not observed snow depth but the proxy h_i constructed from the same observational fields (snow albedo, snow density, sea ice concentration). A.6 confirms the model receives only the snow density time series while the target is a normalized proxy that also depends on C and α. No independent snow-depth observations are used in the paper, and the conclusion explicitly defers validation against independent sources to future work. Consequently the reported 24.4% MSE reduction shows that PhysE-Inv reproduces a self-generated proxy, not that it accurately reconstructs Arctic snow depth.

full rationale

The primary circularity is internal and definitional. Equation 10 defines the 'physics-constrained' prediction as an affine transformation of the model's own direct prediction and its mean; Equation 11 then trains the model to match that direct prediction to the affine-transformed version. The loss can be driven to zero for any prediction by choosing the identity affine parameters, so the physics loss imposes no external hydrostatic-balance constraint. Any improvement in 'physical consistency' attributed to PhysE-Inv is therefore a self-consistency artifact, not agreement with Eq. 4 or Eq. 5. The secondary circularity is the evaluation target: the 'ground truth' is a proxy generated from the same variables the authors describe as available observations, and A.6 lists the input as snow density only while the target is that normalized proxy. The conclusion's future-work sentence admits that independent in-situ/remote-sensing validation has not been performed. Thus the central snow-depth claim reduces to fitting and reproducing a self-generated label; only the narrower claim of proxy-prediction skill relative to baselines retains some internal validity. No load-bearing self-citation chain was found; the circularity is in the paper's own equations and target construction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central evaluation rests on a self-generated proxy target and a fitted three-parameter reconstruction. The model adds no externally validated physical constants or constraints.

free parameters (3)

alpha (hidden physics parameter) = range [-1,1] via sigmoid transform (Eq 7)
Estimated per time step by MLP; used in reconstruction Eq 10; no independent physical meaning.
beta (hidden physics parameter) = range (0,∞) via exp transform (Eq 8)
Estimated per time step; used in Eq 10; fitted to minimize proxy loss.
gamma (hidden physics parameter) = range [-10,10] via tanh transform (Eq 9)
Estimated per time step; additive offset in Eq 10; fitted to data.

assumptions (5)

domain assumption Hydrostatic balance equation relates ice, snow, and water densities to ice thickness/freeboard (Eq 2-4).
Used to construct target proxy; standard sea ice physics but simplified without external validation.
ad hoc to paper Sea ice concentration and snow albedo can serve as proxies for freeboard and snow depth in Eq 5.
Introduced in §4.1.1 without derivation; no empirical calibration to observations.
domain assumption System evolution can be described by linear dynamics in latent space (Kwasniok 2022).
Motivates LSTM/linear inverse head; cited but not verified for this dataset.
domain assumption Surjective many-to-one inverse mapping is appropriate; every latent state maps to at least one physical output.
Used to justify MLP parameter head; plausible but not tested.
domain assumption Gaussian noise augmentation preserves underlying physics for contrastive learning.
Used in PGCL; no analysis of whether arbitrary Gaussian noise keeps physical consistency.

invented entities (1)

Hidden physical parameters α, β, γ
purpose: Recovered per time step to reconstruct snow-depth estimate in Eq 10.
These are not parameters of the hydrostatic balance; they are model-internal variables fit to the proxy target. No independent measurement or falsifiable prediction is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Physics-Encoded Inverse Modeling for Arctic Snow Depth Estimation." pith.science (2026). https://pith.science/paper/CDUWTVLZ

@misc{pith2026260117074,
  author       = {Pith},
  title        = {Pith review of: Physics-Encoded Inverse Modeling for Arctic Snow Depth Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CDUWTVLZ}},
  note         = {Machine review of arXiv:2601.17074}
}

read the original abstract

Accurate estimation of unobserved quantities in time-varying inverse problems remains challenging when observations are sparse and only indirectly related to the target variable. In Arctic climate applications, snow depth over sea ice is not directly available in commonly used reanalysis products and must instead be inferred from related physical and environmental variables. To address this challenge, we introduce Physics-Encoded Inverse Modeling (PhysE-Inv), a framework that combines sequential deep learning with a physics-encoded parameter estimation module for inverse estimation under sparse observational conditions. PhysE-Inv uses an LSTM encoder-decoder to capture temporal dependencies and incorporates contrastive learning to improve the consistency of learned representations. The framework learns structured physics-encoded parameters that are integrated with observational inputs to estimate snow depth proxies. Under the proposed proxy evaluation framework, PhysE-Inv outperforms all evaluated baselines, achieving an average MSE reduction of 24.4\% compared with baseline models and a 17.3\% improvement over the strongest baseline under the parameter estimation setting. These results demonstrate the potential of physics-encoded modeling approaches for estimating unobserved quantities in data-scarce domains where direct observations are limited.

Figures

Figures reproduced from arXiv: 2601.17074 by the authors.

**Figure 2.** Comparison of model performance: box plot of snow depth deviations. The box-and-whisker plots in [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Time series of predicted and estimated mean snow [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Map showing the spatial extent of the central Arctic Ocean, highlighting (orange line) the [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

**Figure 5.** Figure 5: (a) Bijective mapping: unique input-output relationship. (b) Surjective mapping: multiple [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

**Figure 6.** Figure 6: PDF of mean snow depth anomalies for true values and PhysE-Inv predictions. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 2 canonical work pages

[2]

Faroughi, N

S. Faroughi, N. Pawar, C. Fernandes, M. Raissi, S. Das, N. Kalantari, and S. Mahjour. Physics- guided, physics-informed, and physics-encoded neural networks in scientific computing.ArXiv Preprint ArXiv:2211.07377,

arXiv
[4]

URLhttps://rmets.onlinelibrary.wiley.com/ doi/abs/10.1002/qj.3803

doi: https://doi.org/10.1002/qj.3803. URLhttps://rmets.onlinelibrary.wiley.com/ doi/abs/10.1002/qj.3803. M. Innes, A. Edelman, K. Fischer, C. Rackauckas, E. Saba, V . Shah, and W. Tebbutt. A differen- tiable programming system to bridge machine learning and scientific computing.ArXiv Preprint ArXiv:1907.07587,

arXiv 1907
[5]

Karpatne, X

A. Karpatne, X. Jia, and V . Kumar. Knowledge-guided machine learning: Current trends and future prospects.ArXiv Preprint ArXiv:2403.15989,

arXiv
[8]

doi: 10.1038/ s42256-021-00302-5

ISSN 2522-5839. doi: 10.1038/ s42256-021-00302-5. URLhttp://dx.doi.org/10.1038/s42256-021-00302-5. S. Mo, N. Zabaras, X. Shi, and J. Wu. Deep-learning-based inverse modeling approaches: A sub- surface flow example.Water Resources Research, 56:e2019WR026731,

doi:10.1038/s42256-021-00302-5
[12]

org/doi/10.1137/1.9781611977172.79

Available:https://epubs.siam. org/doi/10.1137/1.9781611977172.79. J. Willard, X. Jia, S. Xu, M. Steinbach, and V . Kumar. Integrating scientific knowledge with machine learning for engineering and environmental systems.ACM Computing Surveys, 55(1):1–37,

work page doi:10.1137/1.9781611977172.79
[13]

This Region of Importance, the central Arctic Ocean, is an essential component of the global climate system

A APPENDIX A.1 REGION OFIMPORTANCE The behavior of sea ice and its overlying snow is analyzed within a specific geographical domain. This Region of Importance, the central Arctic Ocean, is an essential component of the global climate system. Figure 4 illustrates the exact spatial extent of this region, which is used for extracting all atmospheric and ocea...

2008
[2019]

C. Rao, H. Sun, and Y . Liu. Hard encoding of physics for learning spatiotemporal dynamics.ArXiv Preprint ArXiv:2105.00557,

arXiv
[2020]

H. Fang, S. Wang, M. Zhou, J. Ding, and P. Xie. Cert: Contrastive self-supervised learning for language understanding.ArXiv Preprint ArXiv:2005.12766,

arXiv 2005

Show all 13 references

[2021]

Kovachki, Z

N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, and A. Anandkumar. Neural operator: Learning maps between function spaces.ArXiv Preprint ArXiv:2108.08481,

arXiv
[2022]

Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, Andr ´as Hor ´anyi, Joaqu ´ın Mu ˜noz- Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, Adrian Simmons, Cor- nel Soci, Saleh Abdalla, Xavier Abellan, Gianpaolo Balsamo, Peter Bechtold, Gionata Biav...

1999
[2023]

Tung Nguyen, Arsh Koneru, Shufan Li, and Aditya Grover

URLhttps://arxiv.org/abs/ 2301.10343. Tung Nguyen, Arsh Koneru, Shufan Li, and Aditya Grover. Physix: A foundation model for physics simulations

arXiv
[2024]

URLhttps://agupubs.onlinelibrary.wiley.com/doi/abs/ 10.1029/2024GL110163

doi: https://doi.org/10.1029/ 2024GL110163. URLhttps://agupubs.onlinelibrary.wiley.com/doi/abs/ 10.1029/2024GL110163. e2024GL110163 2024GL110163. T. Liao and G. Li. Metaheuristic-based inverse design of materials–a survey.Journal Of Materi- omics, 6:414–430,

doi
[2025]

Maziar Raissi, Paris Perdikaris, and George E Karniadakis

URLhttps://arxiv.org/abs/2506.17774. Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational phys...

arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [2]

Faroughi, N

S. Faroughi, N. Pawar, C. Fernandes, M. Raissi, S. Das, N. Kalantari, and S. Mahjour. Physics- guided, physics-informed, and physics-encoded neural networks in scientific computing.ArXiv Preprint ArXiv:2211.07377,

arXiv

[2] [4]

URLhttps://rmets.onlinelibrary.wiley.com/ doi/abs/10.1002/qj.3803

doi: https://doi.org/10.1002/qj.3803. URLhttps://rmets.onlinelibrary.wiley.com/ doi/abs/10.1002/qj.3803. M. Innes, A. Edelman, K. Fischer, C. Rackauckas, E. Saba, V . Shah, and W. Tebbutt. A differen- tiable programming system to bridge machine learning and scientific computing.ArXiv Preprint ArXiv:1907.07587,

arXiv 1907

[3] [5]

Karpatne, X

A. Karpatne, X. Jia, and V . Kumar. Knowledge-guided machine learning: Current trends and future prospects.ArXiv Preprint ArXiv:2403.15989,

arXiv

[4] [8]

doi: 10.1038/ s42256-021-00302-5

ISSN 2522-5839. doi: 10.1038/ s42256-021-00302-5. URLhttp://dx.doi.org/10.1038/s42256-021-00302-5. S. Mo, N. Zabaras, X. Shi, and J. Wu. Deep-learning-based inverse modeling approaches: A sub- surface flow example.Water Resources Research, 56:e2019WR026731,

doi:10.1038/s42256-021-00302-5

[5] [12]

org/doi/10.1137/1.9781611977172.79

Available:https://epubs.siam. org/doi/10.1137/1.9781611977172.79. J. Willard, X. Jia, S. Xu, M. Steinbach, and V . Kumar. Integrating scientific knowledge with machine learning for engineering and environmental systems.ACM Computing Surveys, 55(1):1–37,

work page doi:10.1137/1.9781611977172.79

[6] [13]

This Region of Importance, the central Arctic Ocean, is an essential component of the global climate system

A APPENDIX A.1 REGION OFIMPORTANCE The behavior of sea ice and its overlying snow is analyzed within a specific geographical domain. This Region of Importance, the central Arctic Ocean, is an essential component of the global climate system. Figure 4 illustrates the exact spatial extent of this region, which is used for extracting all atmospheric and ocea...

2008

[7] [2019]

C. Rao, H. Sun, and Y . Liu. Hard encoding of physics for learning spatiotemporal dynamics.ArXiv Preprint ArXiv:2105.00557,

arXiv

[8] [2020]

H. Fang, S. Wang, M. Zhou, J. Ding, and P. Xie. Cert: Contrastive self-supervised learning for language understanding.ArXiv Preprint ArXiv:2005.12766,

arXiv 2005