Hard conservation correctors can hide a degrading model when training autoregressive emulators

John Schreck; William E. Chapman; Yingkai Sha

Hard conservation correctors can hide a degrading model when training autoregressive emulators

T0 review · 0 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A hard water-budget corrector makes raw precipitation amplitude invisible to the training loss, allowing the raw model to degrade while the delivered field stays exactly conservative.

desk verdict A clean, exact argument that hard conservation correctors can hide raw-field drift, backed by a reproducible 2.4% to 24.1% correction growth and a controlled ablation. read the letter →

arxiv 2607.18416 v1 pith:6MFZ377J submitted 2026-07-20 physics.ao-ph

William E. Chapman , John Schreck , Yingkai Sha This is my paper

classification physics.ao-ph

keywords machine-learningweatheremulatorshardconservationcorrectorswater-budgetclosurescaledegeneracyautoregressivetrainingprecipitationbiasdifferentiablecorrectionbudgetdiagnostics

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

This paper shows that a hard global water-budget corrector, applied before the supervised loss during autoregressive training, removes the raw precipitation amplitude from the information the loss can restore. Any uniform rescaling of the raw precipitation field is absorbed by the corrector's rescaling factor, so the corrected output and the loss are unchanged; the amplitude direction lies in the null space of the correction Jacobian. In practice, fine-tuning a global atmosphere emulator this way let the required correction grow from 2.4% to 24.1% over 18 epochs while every delivered field closed the moisture budget to machine precision. Switching to pre-correction supervision plus a soft imbalance penalty returned the correction to under 1% within one epoch, and a 2x2 ablation showed the runaway occurs only when corrected-output supervision is paired with no penalty. The paper's central message is that exact post-correction closure is not a diagnostic of whether the raw model has learned the budget; the raw fields and the correction factor must be tracked.

What carries the argument

The central object is the global water-budget corrector operating on precipitation. It computes the globally integrated total-column-water tendency and evaporation, then rescales the predicted precipitation field by the factor r = (E - T)/P required to close the budget. Because the corrector preserves the normalized spatial pattern P/(1^T P) and only replaces the global amplitude Q, the corrected field is invariant to uniform rescaling of the raw field; the raw-amplitude direction is in the null space of the corrector's Jacobian. That null direction is what carries the argument: it identifies the degree of freedom the supervised loss cannot constrain.

What would settle it

Take the production corrupted configuration (corrected-output supervision, no penalty) and freeze or detach the decoder weights shared between precipitation and the other supervised variables, so that only precipitation receives gradients. If the raw precipitation amplitude still drifts, the null-space mechanism alone is confirmed; if it does not, shared-weight gradients were the constraining force the paper did not isolate. Alternatively, compute the gradient of the corrected-output loss along the raw-amplitude direction J_C(P) P over actual training batches; a nonzero value would refute the

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a scale degeneracy in hard conservation correctors used during training. For the multiplicative precipitation corrector C(P)=Q P / (1^T P), every positive rescaling α of the raw field produces the same corrected field: C(αP)=C(P). Consequently the Jacobian satisfies J_C(P) P = 0, meaning the raw-amplitude direction carries no gradient from a supervised loss evaluated only on C(P). Empirically this degeneracy allowed the raw global precipitation sink to fall roughly 40% below the reference while the corrected field closed the global water budget to machine precision; restoring direct supervision of the pre-correction prediction and adding a penalty o

Load-bearing premise

The load-bearing premise is that the raw-amplitude null direction is not re-coupled to the loss by gradients flowing through shared decoder weights: winds, temperature, and humidity remain supervised, and the paper explicitly states that the authors did not isolate that forcing; if those shared-weight gradients do constrain amplitude in other settings, the failure mode could be configuration-specific rather than a general property of corrected-output supervision.

Editorial extensions

If this is right

If a hard corrector is applied before the supervised loss, the training objective cannot restore the raw property in the corrector's null space; that property must be constrained by other terms.
A closed budget on the delivered field is not evidence that the raw model has learned the budget; a growing correction factor is a warning sign of raw-field deterioration.
Supervising the pre-correction prediction and penalizing the raw budget imbalance stabilizes training while keeping the delivered field exactly conservative.
The same argument applies to any differentiable corrector (mass, energy, surface-pressure) with null directions that overlap weakly supervised model degrees of freedom.
Monitoring tools should report the raw residual and the correction magnitude, not only the post-correction closure.

Reading between the lines

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

This suggests a general pre-training audit: compute the null space of any candidate corrector and check whether it spans directions that are otherwise weakly supervised; if so, plan an explicit constraint on those directions.
Because the drift direction was consistently downward and the paper did not isolate the forcing, a testable extension is to block the shared-weight gradient path between precipitation and other supervised variables; if the drift persists, the null-space mechanism alone is confirmed, and if it disappears, shared-weight gradients were the constraining force.
The same blind spot likely applies to inference-only correctors applied at every autoregressive step: they can hide raw drift over long integrations, so the raw field's amplitude should be logged alongside the correction factor even when no training loss is involved.
This result may extend to non-multiplicative correctors such as projections wherever the correction Jacobian has a null direction; the paper cautions against automatic generalization, but the diagnostic of tracking raw residual and correction magnitude is cheap to apply broadly.

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

You should know two things. First, the paper gives a clean, exact argument for why hard conservation correctors can hide raw-field drift: the water-budget corrector is globally multiplicative, so C(alpha P) = C(P), and a loss on the corrected output has zero gradient along raw amplitude. Second, the empirical demonstration is concrete and reproducible: over 18 epochs the required correction grew from 2.4% to 24.1% while every delivered field closed the moisture budget to machine precision. That is a real warning for the AI weather emulation field.

What is actually new: the scale-invariance identity, connected to a measured training failure, and the general Jacobian null-space diagnostic in Eq. (2). That diagnostic is a useful framing for any differentiable corrector, not just precipitation. The empirical work is solid—the runaway reproduced in a second production run (peaking above 60%), and in ablation cell A, while cells B–D stayed flat. The 2x2 ablation cleanly separates the two interventions: corrected-output supervision and no penalty. The paper is also unusually honest: no forecast skill claims, no free-running stability tests, and it explicitly says the remedy changed two variables at once.

Soft spots, in proportion. The generalizable claim rests on single-seed, single-checkpoint runs, and the ablation spans only four epochs. That is a real limitation for extrapolating to other architectures, but it does not undercut the existence proof—cell A drifts despite shared decoder weights and supervised neighboring fields, so the null direction wins in practice. The remedy penalizes (r-1)^2 directly, so the headline "correction back under 1%" is partly by construction. The paper acknowledges this and offers the raw sink recovery as a partly independent check; that is fair, though the sink depends on unbiased evaporation and storage predictions. The ACE2 comparison is careful not to claim ACE2 drifts, only that its diagnostics would miss it—fine.

Who this is for: anyone training or evaluating AI weather/climate emulators with hard correctors, and people designing conservation diagnostics. It deserves a serious referee. The mathematical identity is exact, the empirical trajectory is convincing, and the recommendations—track raw residual and correction magnitude—are practical.

Recommendation: send it to peer review. Ask for multi-seed reproduction and at least one long free-running diagnostic, but the core contribution is publishable as is.

Referee Report

0 major / 5 minor

Summary. The paper documents a failure mode in training autoregressive weather emulators with hard conservation correctors. When the supervised loss is evaluated on the corrected output, a global multiplicative water-budget corrector introduces a null direction in the loss corresponding to uniform rescaling of raw precipitation amplitude (Eq. 1). The raw amplitude therefore receives no direct restoring gradient from the corrected-output term. Fine-tuning experiments on the CAMulator show the required correction growing from 2.4% to 24.1% over 18 epochs while every delivered field closes the moisture budget to machine precision. A second run reproduces the drift, and a 2x2 ablation isolates the failure to the combination of corrected-output supervision with no imbalance penalty. Switching to pre-correction supervision with a soft imbalance penalty (Eq. 7) returns the required correction to below 1% within one epoch. The paper argues that post-correction budget closure is not an independent diagnostic and recommends tracking raw fields and the correction magnitude.

Significance. If the result holds, this is an important cautionary lesson for the growing practice of embedding physical constraints in ML emulators. The central null-space argument is clean and exact, and the numerical verification of C(alpha P) = C(P) is convincing. The empirical demonstration is strengthened by a second production run, a controlled 2x2 ablation, and a third occurrence of the runaway in ablation cell A. The paper is appropriately modest in scope: it explicitly acknowledges that shared-weight gradients were not isolated, that the ablation is from a single checkpoint, that no free-running skill was measured, and that the remedy's penalty directly targets (r-1)^2. These limitations affect breadth of generalization but do not undermine the existence proof that a hard corrector can conceal raw-field degradation. The paper ships code, configurations, and figure-generation scripts, which supports reproducibility.

minor comments (5)

[General formatting] Throughout the manuscript there are missing spaces and typographical glitches: e.g., 'factorr, wherer = 1' in Section 2.2, 'controlled2× 2ablation' in Section 2.4, and 'Fig. 1b'/'Fig. 1c' spacing. A careful proofreading pass is needed.
[Appendix, Eq. (3)] The latitude weight is written as L_lat_i = cos(phi_i)^p / cos(phi)^p, but the exponent p is never defined. Please provide the definition or a reference, as this is part of the loss used in the experiments.
[Section 3.2 and Discussion, Eq. (2)] The general Jacobian criterion is stated as making a direction 'invisible to that objective.' This is exact for the corrected-output term alone, but in a multi-output network a change in raw precipitation amplitude can be coupled to other, still-supervised outputs through shared weights. The paper acknowledges this in Section 3.2 ('We did not isolate that forcing'), but a one-sentence qualification near Eq. (2) would prevent over-reading of the general statement.
[Section 3.4] The ablation is described as single runs from one checkpoint, and the paper notes that cells B-D are not distinguishable from one another at this sample size. That is transparent. Consider stating even more prominently that the ablation establishes the short-term mechanism only, not long-term stable training, to avoid the cells B-D being interpreted as performance validation.
[Section 3.3] The sentence 'Because the soft penalty targets (r-1)^2 directly, the collapse of the required correction is necessary but not sufficient evidence of an improved model' is commendably honest. The independent support from the raw sink recovery would be stronger if the figure included a quantitative comparison of the raw sink to the CAM6 reference with uncertainty, but this is optional.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Eq. (1) is an exact consequence of the defined corrector, and the empirical drift is reproduced and ablated, with limitations disclosed.

full rationale

The central claim is the scale-degeneracy identity C(αP)=C(P) and J_C(P)P=0 (Eq. 1), which the paper derives directly from the stated correction operator C(P)=Q P/1^T P; it is a mathematical consequence of the definition, not a fitted result. The empirical evidence for the failure mode consists of observed training trajectories (correction growing 2.4%–24.1%, a second run above 60%, and ablation cell A drifting while B–D stay flat), not predictions obtained from fitted parameters. The remedy adds a penalty λ_w(r−1)² that directly targets the measured correction; the paper explicitly flags this ('necessary but not sufficient') and supplements it with the ablation and raw-sink recovery, while noting that raw-field forecast skill was not measured. The 2×2 ablation is a controlled comparison, not a circular restatement: cell C (pre-correction supervision, no penalty) and cell B (corrected output, penalty) each suppress the runaway, matching the proposed mechanism. The raw-sink recovery in the production remedy is partly by construction because the stable objective also directly supervises the pre-correction precipitation; the paper's 'partly independent' phrasing is generous, but that auxiliary indicator is not load-bearing for the main conclusion. Self-citations (Sha et al. 2025, Chapman et al. 2025) are used for motivation, setup context, and a methodological choice; the null-space proof and ablation do not rest on them. Stated limitations (shared-weight gradients not isolated, single checkpoint, no free-running skill, two-variable change) are honest scope restrictions rather than hidden inputs. No self-definitional, fitted-input-as-prediction, or self-citation-load-bearing step is present.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper contributes an identity and an empirical training study, not a new physical entity. The ledger records the corrector definition, the domain assumptions needed to generalize from one emulator, and the single tuned penalty weight. No new forces, particles, dimensions, or conserved quantities are introduced.

free parameters (1)

lambda_w (imbalance penalty weight) = 0.1
Chosen by hand for the stable runs in Eq. (7). The ablation shows pre-correction supervision alone also prevents the runaway, so this penalty weight is a tuning choice rather than a fitted prediction.

assumptions (3)

domain assumption The water-budget corrector is a globally uniform multiplicative rescaling of the precipitation field: C(P) = Q P / (1^T P), preserving the spatial pattern.
Defined in Section 2.2 and Appendix Eq. (5). The null-space argument is specific to this operator, which is the operator used in CREDIT and ACE2.
domain assumption Gradients from other supervised fields sharing precipitation decoder weights do not fully constrain raw precipitation amplitude during fine-tuning.
Needed to explain why the null direction is not filled by correlated gradients. Section 3.2 identifies this as plausible but says the forcing was not isolated.
domain assumption The CAM6 reference data are representative of a closed water budget, with mean correction 0.024% and per-step spread 2.22%.
Used in Section 2.3 to establish that a persistent order-10% correction is pathological rather than a target-data artifact. Verified for one year of CAM6 output, not for the whole training set.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hard conservation correctors can hide a degrading model when training autoregressive emulators." pith.science (2026). https://pith.science/paper/6MFZ377J

@misc{pith2026260718416,
  author       = {Pith},
  title        = {Pith review of: Hard conservation correctors can hide a degrading model when training autoregressive emulators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6MFZ377J}},
  note         = {Machine review of arXiv:2607.18416}
}

read the original abstract

AI weather and climate emulators increasingly incorporate physical principles into their formulation. One approach is to apply hard correctors that modify network outputs so that global mass, water, or energy budgets close. Prior work introduced such training-time correctors in the CREDIT framework and reported reduced precipitation bias and improved stability. Motivated by those results, we fine-tuned a global atmosphere emulator with a water-budget corrector, using the corrected prediction in the supervised loss and evaluating through post-correction budget closure. By that measure, training appeared successful. Every delivered field closed the moisture budget to machine precision. However, raw precipitation developed a growing global low bias over 18 training epochs, while the required correction increased from about 2% to roughly 24%. The cause is a scale degeneracy. A uniform change in raw precipitation amplitude is offset by a compensating change in the correction factor, leaving the corrected field, and therefore the supervised loss, unchanged. This invariance removes the restoring force on raw precipitation amplitude, allowing other training pressures to drive drift. Two changes recovered stable behavior. We supervised the pre-correction prediction and penalized its raw budget imbalance, while the hard correction remained in place for the delivered field. The required correction returned to less than 1% within the next epoch. A controlled 2x2 ablation showed that the runaway occurred only when corrected-output supervision was combined with no imbalance penalty. Exact post-correction closure therefore says little about whether the raw model has learned the budget. When a corrector removes information from the loss, the raw fields and the applied correction need to be tracked.

Figures

Figures reproduced from arXiv: 2607.18416 by the authors.

**Figure 1.** Schematic of the three water-budget training configurations. (a) The raw network prediction is supervised directly, with no budget correction. (b) The water-budget corrector is applied before the supervised loss, so training is based on the corrected prediction. (c) The raw prediction is supervised and penalized for budget imbalance, while the corrector is retained for the delivered output. In both corrected configu… view at source ↗

**Figure 2.** Evolution of the water-budget correction during fine-tuning. (a) The required correction grows to about 24% when the supervised loss is evaluated on the corrected output, then returns to the CAM6 reference range after switching to pre-correction supervision with an imbalance penalty. The dashed line marks this change in training configuration. (b) Distributions of the required correction for CAM6 and the two trainin… view at source ↗

**Figure 3.** Results of the 2 × 2 ablation crossing the supervised target, corrected or pre-correction, with an imbalance-penalty weight of 0 or 0.1. (a) The required water-budget correction grows only for corrected-output supervision without a penalty (cell A); the other three configurations remain near the CAM6 reference range. (b) The raw global precipitation sink for the same experiments. Cell A develops a large low bias, wh… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 2 canonical work pages

[3]

Chapman, John S

William E. Chapman, John S. Schreck, Yingkai Sha, David John Gagne, II, Dhamma Kimpara, Laure Zanna, Kirsten J. Mayer, and Judith Berner. CAMulator: Fast emulation of the community atmosphere model.arXiv preprint arXiv:2504.06007,

arXiv
[7]

doi: 10.1038/s42254-021-00314-5. Dmitrii Kochkov, Janni Yuval, Ian Langmore, Peter Norgaard, Jamie Smith, Griffin Mooers, Milan Klöwer, James Lottes, Stephan Rasp, Peter Düben, Sam Hatfield, Peter Battaglia, Alvaro Sanchez- Gonzalez, Matthew Willson, Michael P. Brenner, and Stephan Hoyer. Neural general circulation models for weather and climate.Nature, 6...

doi:10.1038/s42254-021-00314-5
[8]

doi: 10.1038/s41586-024-077 44-y. Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, Alexander Merose, Stephan Hoyer, George Holland, Oriol Vinyals, Jacklynn Stott, Alexander Pritzel, Shakir Mohamed, and Peter Battaglia. Learning skillful medium-range...

work page doi:10.1038/s41586-024-077
[9]

Maziar Raissi, Paris Perdikaris, and George Em Karniadakis

doi: 10.1126/science.adi2336. Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 378:686–707,

doi:10.1126/science.adi2336
[11]

Christopher Subich, Syed Zahid Husain, Leo Separovic, and Jing Yang

doi: 10.1029/2025MS005138. Christopher Subich, Syed Zahid Husain, Leo Separovic, and Jing Yang. Fixing the double penalty in data-driven weather forecasting through a modified spherical harmonic loss function. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 ofProceedings of Machine Learning Research, pages 57191–57211. PMLR,

doi:10.1029/2025ms005138
[12]

Oliver Watt-Meyer, Gideon Dresdner, Jeremy McGibbon, Spencer K

doi: 10.1038/s42005-025-02329-1. Oliver Watt-Meyer, Gideon Dresdner, Jeremy McGibbon, Spencer K. Clark, Brian Henn, James Duncan, Noah D. Brenowitz, Karthik Kashinath, Michael S. Pritchard, Boris Bonev, Matthew E. Peters, and Christopher S. Bretherton. ACE: A fast, skillful learned global atmospheric model for climate prediction.arXiv preprint arXiv:2310.02074,

arXiv
[13]

Accepted at the NeurIPS 2023 Workshop on Tackling Climate Change with Machine Learning

doi: 10.48550/arXiv.2310.02074. Accepted at the NeurIPS 2023 Workshop on Tackling Climate Change with Machine Learning. Oliver Watt-Meyer, Brian Henn, Jeremy McGibbon, Spencer K. Clark, Anna Kwa, W. Andre Perkins, Elynn Wu, Lucas Harris, and Christopher S. Bretherton. ACE2: Accurately learning subseasonal to decadal atmospheric variability and forced resp...

arXiv doi:10.48550/arxiv.2310.02074 2023
[14]

Francesco Zanetta, Daniele Nerini, Tom Beucler, and Mark A

doi: 10.1038/s41612-025-01090-0. Francesco Zanetta, Daniele Nerini, Tom Beucler, and Mark A. Liniger. Physics-constrained deep learning postprocessing of temperature and humidity.Artificial Intelligence for the Earth Systems, 2(4):e220089,

doi:10.1038/s41612-025-01090-0

Show all 15 references

[15]

doi: 10.1175/AIES-D-22-0089.1. 12

doi
[2009]

10 Chapman et al

doi: 10.1175/2009WAF2222269.1. 10 Chapman et al. Hard conservation correctors in autoregressive emulators William Gregory, Mitchell Bushuk, James Duncan, Elynn Wu, Adam Subel, Spencer K Clark, Bill Hurlin, Oliver Watt-Meyer, Alistair Adcroft, Chris Bretherton, et al. Floenet: ...

arXiv
[2019]

2018.10.045

doi: 10.1016/j.jcp. 2018.10.045. John S Schreck, William E Chapman, Charlie Becker, David John Gagne II, Dhamma Kimpara, Nihanth Cherukuru, Judith Berner, Kirsten J Mayer, and Negin Sobhani. Controllable probabilistic forecasting with stochastic decomposition layers.arXiv prep...

2018
[2021]

Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian

doi: 10.1103/PhysRevLett.126.098302. Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium- range global weather forecasting with 3D neural networks.Nature, 619(7970):533–538,

doi
[2023]

Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar

doi: 10.1038/s41586-023-06185-3. Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar. Spherical fourier neural operators: Learning stable dynamics on the sphere. InProceedings of the 40th International Confere...

doi
[2024]

George Em Karniadakis, Ioannis G

doi: 10.1029/2023MS004021. George Em Karniadakis, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3:422–440,

doi
[2025]

Eric Gilleland, David Ahijevych, Barbara G

doi: 10.48550/arXiv.2504.06007. Eric Gilleland, David Ahijevych, Barbara G. Brown, Barbara Casati, and Elizabeth E. Ebert. Intercomparison of spatial forecast verification methods.Weather and Forecasting, 24(5):1416– 1430,

doi arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [3]

Chapman, John S

William E. Chapman, John S. Schreck, Yingkai Sha, David John Gagne, II, Dhamma Kimpara, Laure Zanna, Kirsten J. Mayer, and Judith Berner. CAMulator: Fast emulation of the community atmosphere model.arXiv preprint arXiv:2504.06007,

arXiv

[2] [7]

doi: 10.1038/s42254-021-00314-5. Dmitrii Kochkov, Janni Yuval, Ian Langmore, Peter Norgaard, Jamie Smith, Griffin Mooers, Milan Klöwer, James Lottes, Stephan Rasp, Peter Düben, Sam Hatfield, Peter Battaglia, Alvaro Sanchez- Gonzalez, Matthew Willson, Michael P. Brenner, and Stephan Hoyer. Neural general circulation models for weather and climate.Nature, 6...

doi:10.1038/s42254-021-00314-5

[3] [8]

doi: 10.1038/s41586-024-077 44-y. Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, Alexander Merose, Stephan Hoyer, George Holland, Oriol Vinyals, Jacklynn Stott, Alexander Pritzel, Shakir Mohamed, and Peter Battaglia. Learning skillful medium-range...

work page doi:10.1038/s41586-024-077

[4] [9]

Maziar Raissi, Paris Perdikaris, and George Em Karniadakis

doi: 10.1126/science.adi2336. Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 378:686–707,

doi:10.1126/science.adi2336

[5] [11]

Christopher Subich, Syed Zahid Husain, Leo Separovic, and Jing Yang

doi: 10.1029/2025MS005138. Christopher Subich, Syed Zahid Husain, Leo Separovic, and Jing Yang. Fixing the double penalty in data-driven weather forecasting through a modified spherical harmonic loss function. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 ofProceedings of Machine Learning Research, pages 57191–57211. PMLR,

doi:10.1029/2025ms005138

[6] [12]

Oliver Watt-Meyer, Gideon Dresdner, Jeremy McGibbon, Spencer K

doi: 10.1038/s42005-025-02329-1. Oliver Watt-Meyer, Gideon Dresdner, Jeremy McGibbon, Spencer K. Clark, Brian Henn, James Duncan, Noah D. Brenowitz, Karthik Kashinath, Michael S. Pritchard, Boris Bonev, Matthew E. Peters, and Christopher S. Bretherton. ACE: A fast, skillful learned global atmospheric model for climate prediction.arXiv preprint arXiv:2310.02074,

arXiv

[7] [13]

Accepted at the NeurIPS 2023 Workshop on Tackling Climate Change with Machine Learning

doi: 10.48550/arXiv.2310.02074. Accepted at the NeurIPS 2023 Workshop on Tackling Climate Change with Machine Learning. Oliver Watt-Meyer, Brian Henn, Jeremy McGibbon, Spencer K. Clark, Anna Kwa, W. Andre Perkins, Elynn Wu, Lucas Harris, and Christopher S. Bretherton. ACE2: Accurately learning subseasonal to decadal atmospheric variability and forced resp...

arXiv doi:10.48550/arxiv.2310.02074 2023

[8] [14]

Francesco Zanetta, Daniele Nerini, Tom Beucler, and Mark A

doi: 10.1038/s41612-025-01090-0. Francesco Zanetta, Daniele Nerini, Tom Beucler, and Mark A. Liniger. Physics-constrained deep learning postprocessing of temperature and humidity.Artificial Intelligence for the Earth Systems, 2(4):e220089,

doi:10.1038/s41612-025-01090-0