pith. sign in

arxiv: 2606.02610 · v2 · pith:XNYOMHAOnew · submitted 2026-05-24 · 💻 cs.CE · cs.AI· cs.LG· physics.ao-ph

Samudra 2: Scaling Ocean Emulators across Resolutions

Yuan Yuan , Jesse Rusak , Alexander Merose , Adam Subel , Pavel Perezhogin , Alistair Adcroft , Carlos Fernandez-Granda , Laure Zanna This is my paper

Pith reviewed 2026-06-29 23:30 UTC · model grok-4.3

classification 💻 cs.CE cs.AIcs.LGphysics.ao-ph
keywords ocean emulatorneural networkU-Netautoregressive rolloutocean general circulation modelclimate modelingmesoscale eddiesresolution scaling
0
0 comments X

The pith

Samudra 2 uses a wider U-Net and dynamic loss reweighting to scale ocean emulators to finer resolutions with stable multi-year rollouts.

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

Samudra 2 advances the original neural ocean emulator by widening the U-Net backbone, adopting modified ConvNeXt-style blocks with a reduced expansion factor, and introducing a dynamic loss that reweights channels by their prediction errors. These changes raise upper-ocean global-mean temperature R² from 0.56 to 0.87 at 1° resolution and cut deep-ocean temperature error by roughly seven times. The identical architecture then produces stable autoregressive rollouts at 1/2° and 1/4° resolution over about eight years while recovering mesoscale eddies and sharp western boundary currents. The work addresses prior failure modes of variance collapse and imprinting artifacts. Faster single-GPU execution opens the door to larger ensembles for sea-level, heat-uptake, and variability studies.

Core claim

Samudra 2 demonstrates that a wider U-Net backbone with modified ConvNeXt-style blocks, reduced internal expansion factor, and dynamic loss reweighting that strengthens gradients for slow-evolving deep-ocean fields enables an autoregressive neural emulator to improve temperature predictions at 1° resolution and to scale stably to 1/2° and 1/4° resolutions, sustaining approximately 8-year global rollouts that recover mesoscale eddies and western boundary currents.

What carries the argument

Wider U-Net backbone with modified ConvNeXt-style blocks, reduced block-internal expansion factor, and dynamic loss reweighting that strengthens gradients for slow-evolving fields.

If this is right

  • Single-GPU execution enables larger ensembles for sea-level projections, ocean heat uptake, and climate variability studies.
  • The same architecture works across 1°, 1/2°, and 1/4° resolutions without resolution-specific retraining.
  • Recovery of mesoscale eddies and sharp western boundary currents improves representation of ocean dynamics at finer grids.
  • Stable multi-year rollouts become feasible for climate studies that previously required slower traditional models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Faster emulation could lower the barrier to exploring many forcing scenarios in ensemble climate projections.
  • Stable scaling across resolutions suggests the modifications may support even higher resolutions if additional compute is available.
  • Public code and checkpoints allow direct testing on additional ocean variables or regional domains.

Load-bearing premise

The modifications to the U-Net backbone and the dynamic loss reweighting are the primary drivers of the accuracy gains and the ability to scale stably to finer resolutions.

What would settle it

An 8-year autoregressive rollout at 1/4° resolution that exhibits variance collapse or imprinting artifacts would show the scaling claim does not hold.

Figures

Figures reproduced from arXiv: 2606.02610 by Adam Subel, Alexander Merose, Alistair Adcroft, Carlos Fernandez-Granda, Jesse Rusak, Laure Zanna, Pavel Perezhogin, Yuan Yuan.

Figure 1
Figure 1. Figure 1: Surface kinetic energy in the Gulf Stream region from GFDL OM4 (top) and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training (a) vs. inference (b) rollout. During training, the emulator is unrolled for [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The six evaluation regions: four western boundary current regions (Gulf Stream, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview comparison of Samudra and Samudra 2 at 1 [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Temporal variance of temperature across depth layers (columns) and models [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Temporal variance of salinity across depth layers and models. Samudra shows [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Temporal variance of EKE across depth layers (columns) and models (rows). [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Deseasonalized temperature anomaly snapshot at 2022-09-30, near the end of the [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Isotropic power spectra for six ocean regions (Gulf Stream, Kuroshio, Agulhas, [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Detrended global mean temperature time series for three depth ranges (Upper [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Deseasonalized zonal velocity (u) anomaly snapshot at 2022-09-30, near the end [PITH_FULL_IMAGE:figures/full_fig_p030_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Deseasonalized EKE snapshot at 2022-09-30, near the end of the 8-year rollout, [PITH_FULL_IMAGE:figures/full_fig_p031_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Temporal EKE power spectra for six ocean regions. Each panel shows the OM4 [PITH_FULL_IMAGE:figures/full_fig_p032_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Temporal temperature power spectra for six ocean regions. Each panel shows [PITH_FULL_IMAGE:figures/full_fig_p032_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Autocorrelation function (ACF) of surface velocity anomalies across six ocean [PITH_FULL_IMAGE:figures/full_fig_p033_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Autocorrelation function of SSH anomalies across six ocean regions. All emula [PITH_FULL_IMAGE:figures/full_fig_p033_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Detrended global mean temperature time series for three depth ranges for Samu [PITH_FULL_IMAGE:figures/full_fig_p034_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Detrended global mean salinity time series for three depth ranges for Samudra [PITH_FULL_IMAGE:figures/full_fig_p034_18.png] view at source ↗
read the original abstract

Ocean general circulation models (OGCMs) are essential to climate science but computationally expensive, limiting ensemble size and forcing scenarios. Neural emulators promise orders-of-magnitude speedups, yet existing ocean emulators have not combined fine spatial resolution with multi-year autoregressive rollouts. Samudra, the first autoregressive neural ocean emulator to produce multi-decade global rollouts, is limited to $1^\circ$ resolution and exhibits two long-horizon failure modes: \emph{variance collapse}, the loss of temporal variability, and \emph{imprinting artifacts}, in which velocity patterns leak into deep-ocean fields. We present Samudra 2, which introduces a wider U-Net backbone with modified ConvNeXt-style blocks and a reduced block-internal expansion factor, together with a dynamic loss that reweights output channels according to their prediction errors, strengthening gradients for slow-evolving deep-ocean fields. At $1^\circ$, Samudra 2 increases upper-ocean global-mean temperature $R^2$ from 0.56 to 0.87 and reduces deep-ocean temperature error by roughly sevenfold. The same architecture scales to $1/2^\circ$ and $1/4^\circ$ over approximately 8-year autoregressive rollouts, recovering mesoscale eddies and sharp western boundary currents. Running on a single GPU, Samudra 2 enables larger ensembles for sea-level projections, ocean heat uptake, and climate variability studies. All artifacts are publicly available: project page, code, checkpoints, documentation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper presents Samudra 2, an improved neural ocean emulator based on a wider U-Net backbone using modified ConvNeXt-style blocks with reduced expansion factor, plus a dynamic loss reweighting scheme that adjusts channel weights by prediction error. It claims this addresses variance collapse and imprinting, raising upper-ocean global-mean temperature R² from 0.56 to 0.87 at 1° while reducing deep-ocean temperature error by roughly sevenfold, and enabling stable scaling to ½° and ¼° resolutions over ~8-year autoregressive rollouts that recover mesoscale eddies and western boundary currents. All code, checkpoints, and documentation are released publicly.

Significance. If the quantitative gains and scaling behavior are confirmed, the work would be significant for climate applications by enabling single-GPU high-resolution ocean emulations that support larger ensembles for sea-level, heat uptake, and variability studies. The public release of artifacts is a clear strength that aids reproducibility and follow-on work.

major comments (2)
  1. [Results section] Results section: The central performance claims (R² increase from 0.56 to 0.87 and sevenfold deep-ocean error reduction at 1°) are reported without error bars, explicit training/validation splits, or controls for post-hoc tuning, which are required to substantiate the quantitative improvements over the prior Samudra model.
  2. [Methods and Results sections] Methods and Results sections: The manuscript attributes the gains and successful scaling to the wider U-Net (modified ConvNeXt blocks, reduced expansion factor) and dynamic loss reweighting, but supplies no ablation experiments isolating these changes from capacity increases or training schedule variations; this leaves the causal link to the listed modifications unverified and is load-bearing for the scaling claim.
minor comments (1)
  1. [Abstract] The abstract states 'approximately 8-year autoregressive rollouts' at finer resolutions; the main text should report the precise rollout lengths, ensemble sizes, and any resolution-specific training details for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the referee's constructive comments. We address each major comment point by point below.

read point-by-point responses

  1. Referee: [Results section] Results section: The central performance claims (R² increase from 0.56 to 0.87 and sevenfold deep-ocean error reduction at 1°) are reported without error bars, explicit training/validation splits, or controls for post-hoc tuning, which are required to substantiate the quantitative improvements over the prior Samudra model.

    Authors: We agree that the quantitative claims would be strengthened by error bars, explicit split details, and tuning controls. The training/validation splits follow the protocol from the original Samudra paper and are described in Section 3, but we will expand this with a dedicated table for clarity. We will retrain with three random seeds to report means and standard deviations for the key R² and error metrics. We will also add a Methods paragraph detailing the hyperparameter search and validation-based selection process. These changes will be incorporated in the revision. revision: yes

  2. Referee: [Methods and Results sections] Methods and Results sections: The manuscript attributes the gains and successful scaling to the wider U-Net (modified ConvNeXt blocks, reduced expansion factor) and dynamic loss reweighting, but supplies no ablation experiments isolating these changes from capacity increases or training schedule variations; this leaves the causal link to the listed modifications unverified and is load-bearing for the scaling claim.

    Authors: We acknowledge that the lack of ablations leaves the individual contributions less isolated than ideal. The reported gains are shown via direct comparison to the prior Samudra baseline, which used a narrower architecture and static loss. In revision we will add targeted ablations at 1° resolution comparing (i) new backbone with original loss and (ii) original backbone with dynamic loss. Full factorial ablations across resolutions are computationally prohibitive, but these experiments will better support the attribution while we note the combined effect enables the observed scaling. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical metrics are direct measurements on held-out data

full rationale

The paper reports measured R² values (0.56→0.87), error reductions, and rollout stability from training a modified U-Net on ocean simulation data and evaluating autoregressive performance on held-out periods. These quantities are obtained by direct comparison to reference OGCM output and do not reduce, by any equation or self-citation chain inside the paper, to quantities defined by the model's own fitted parameters or prior Samudra results. The reference to the original Samudra work supplies only baseline context and is not invoked to justify uniqueness or to derive the new performance numbers. No self-definitional, fitted-input-called-prediction, or ansatz-smuggling patterns appear in the derivation of the reported gains.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on empirical training and evaluation of a neural network on ocean simulation data. No free parameters are introduced that are fitted to produce the headline metrics, no new axioms are invoked, and no new physical entities are postulated.

pith-pipeline@v0.9.1-grok · 5841 in / 1193 out tokens · 39019 ms · 2026-06-29T23:30:47.452293+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 25 canonical work pages · 4 internal anchors

  1. [1]

    Early View

    doi: 10.1029/2025MS004991. Early View. Boris Bonev, Thorsten Kurth, et al. FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale.arXiv preprint arXiv:2507.12144,

    work page doi:10.1029/2025ms004991

  2. [2]

    20 Samudra 2 Ashesh Chattopadhyay, Y

    doi: 10.48550/arXiv.2507.12144. 20 Samudra 2 Ashesh Chattopadhyay, Y. Qiang Sun, and Pedram Hassanzadeh. Challenges of learning multi-scale dynamics with AI weather models: Implications for stability and one solution. arXiv preprint arXiv:2304.07029,

    work page doi:10.48550/arxiv.2507.12144

  3. [3]

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich

    doi: 10.48550/arXiv.2304.07029. Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. GradNorm: Gra- dient normalization for adaptive loss balancing in deep multitask networks. InProceedings of the 35th International Conference on Machine Learning (ICML), pages 794–803,

    work page doi:10.48550/arxiv.2304.07029

  4. [4]

    Surya Dheeshjith, Adam Subel, and others

    doi: 10.1038/s41467-025-57389-2. Surya Dheeshjith, Adam Subel, and others. Transfer learning for emulating ocean climate variability across CO2 forcing.arXiv preprint arXiv:2405.18585,

    work page doi:10.1038/s41467-025-57389-2

  5. [5]

    doi: 10.1029/2024GL114318. James P. C. Duncan, Elynn Wu, et al. SamudrACE: Fast and accurate coupled climate modeling with 3D ocean and atmosphere emulators.arXiv preprint arXiv:2509.12490,

    work page doi:10.1029/2024gl114318

  6. [6]

    An evolving coupled model intercomparison project phase 7 (cmip7) and fast track in support of future climate assessment.EGUsphere, 2024:1–51,

    John Patrick Dunne et al. An evolving coupled model intercomparison project phase 7 (cmip7) and fast track in support of future climate assessment.EGUsphere, 2024:1–51,

    2024

  7. [7]

    Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6): Experimental design and organization.Geoscientific Model Development, 9(5):1937–1958,

    Veronika Eyring, Sandrine Bony, et al. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6): Experimental design and organization.Geoscientific Model Development, 9(5):1937–1958,

    1937

  8. [8]

    Hugo Frezat, Julien Le Sommer, Ronan Fablet, Guillaume Balarac, and Redouane Lguen- sat

    doi: 10.3389/fmars.2019.00065. Hugo Frezat, Julien Le Sommer, Ronan Fablet, Guillaume Balarac, and Redouane Lguen- sat. A posteriori learning for quasi-geostrophic turbulence parametrization.Journal of Advances in Modeling Earth Systems, 14(11):e2022MS003124,

    work page doi:10.3389/fmars.2019.00065 2019

  9. [9]

    Piyush Garg, Diana R

    doi: 10.1029/ 2022MS003124. Piyush Garg, Diana R. Gergel, Andrew E. Shao, and Galen J. Yacalis. The recipe matters more than the kitchen: Mathematical foundations of the AI weather prediction pipeline. arXiv preprint arXiv:2604.01215,

    work page arXiv

  10. [10]

    Arthur P

    doi: 10.1016/j.jcp.2022.111090. Arthur P. Guillaumin and Laure Zanna. Stochastic-deep learning parameterization of ocean momentum forcing.Journal of Advances in Modeling Earth Systems, 13(9): e2021MS002534,

    work page doi:10.1016/j.jcp.2022.111090 2022

  11. [11]

    Zijie Guo, Pumeng Lyu, et al

    doi: 10.1029/2021MS002534. Zijie Guo, Pumeng Lyu, et al. Data-driven global ocean modeling for seasonal to decadal prediction.Science Advances, 11(33):eadu2488,

    work page doi:10.1029/2021ms002534

  12. [12]

    Mastering Diverse Domains through World Models

    doi: 10.1126/sciadv.adu2488. Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models.arXiv preprint arXiv:2301.04104,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1126/sciadv.adu2488

  13. [13]

    doi: 10.1016/j.ocemod.2013. 08.007. Yoo-Geun Ham, Jeong-Hwan Kim, and Jing-Jia Luo. Deep learning for multi-year ENSO forecasts.Nature, 573:568–572,

    work page doi:10.1016/j.ocemod.2013 2013

  14. [14]

    Forecasting global weather with graph neural networks.arXiv preprint arXiv:2202.07575,

    Ryan Keisler. Forecasting global weather with graph neural networks.arXiv preprint arXiv:2202.07575,

    work page arXiv

  15. [15]

    Deep multi-scale video prediction beyond mean square error

    doi: 10.1029/2024MS004883. Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond mean square error.arXiv preprint arXiv:1511.05440,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1029/2024ms004883

  16. [16]

    FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators

    Jaideep Pathak, Shashank Subramanian, et al. FourCastNet: A global data-driven high- resolution weather forecasting model.arXiv preprint arXiv:2202.11214,

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    doi: 10.1029/2025GL117046. Boris T. Polyak and Anatoli B. Juditsky. Acceleration of stochastic approximation by averaging.SIAM Journal on Control and Optimization, 30(4):838–855,

    work page doi:10.1029/2025gl117046

  18. [18]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox

    doi: 10.1038/s41586-021-03854-z. Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InMedical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234–241,

    work page doi:10.1038/s41586-021-03854-z

  19. [19]

    Exploring the design space of deep-learning-based weather forecasting systems.arXiv preprint arXiv:2410.07472,

    Shoaib Ahmed Siddiqui et al. Exploring the design space of deep-learning-based weather forecasting systems.arXiv preprint arXiv:2410.07472,

    work page arXiv

  20. [20]

    doi: 10.48550/arXiv.2410. 07472. Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization.arXiv preprint arXiv:1607.08022,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410

  21. [21]

    Applications of deep learning parameterization of ocean momentum forcing.arXiv preprint arXiv:2406.03659,

    Guosong Wang, Min Hou, et al. Applications of deep learning parameterization of ocean momentum forcing.arXiv preprint arXiv:2406.03659,

    work page arXiv

  22. [22]

    doi: 10.48550/arXiv.2406. 03659. Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks.SIAM Journal on Scientific Computing, 43(5):A3055–A3081,

    work page doi:10.48550/arxiv.2406

  23. [23]

    23 Yuan et al

    doi: 10.1137/20M1318043. 23 Yuan et al. Oliver Watt-Meyer, Gideon Dresdner, et al. ACE: A fast, skillful learned global atmospheric model for climate prediction.arXiv preprint arXiv:2310.02074,

    work page doi:10.1137/20m1318043

  24. [24]

    Sergey Zagoruyko and Nikos Komodakis

    doi: 10.1029/2023MS003915. Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. InProceedings of the British Machine Vision Conference (BMVC),

    work page doi:10.1029/2023ms003915

  25. [25]

    Laure Zanna and Thomas Bolton

    doi: 10.5244/C.30.87. Laure Zanna and Thomas Bolton. Data-driven equation discovery of ocean mesoscale closures.Geophysical Research Letters, 47(17):e2020GL088376,

    work page doi:10.5244/c.30.87

  26. [26]

    24 Samudra 2 Appendix A

    doi: 10.1093/nsr/ nwac044. 24 Samudra 2 Appendix A. Implementation Details This appendix summarizes implementation details, derived quantities, and evaluation met- rics. Unless otherwise noted, all quantities are computed on regular latitude-longitude grids after conservative regridding from the native OM4 tripolar output withxesmf. A.1 Training Setup and...

    work page doi:10.1093/nsr/

  27. [27]

    No weight decay or dropout is applied

    over 70 epochs (no warmup), and gradient clipping at 1.0. No weight decay or dropout is applied. The final model is the last-epoch checkpoint rather than the one with the lowest validation loss, since single-step validation loss is not fully predictive of long-horizon rollout performance; in practice, we observed no significant overfitting over the 70-epo...

    1992

  28. [28]

    These changes maintain training dynamics comparable to the 1◦ and 1/2◦ configurations

    with instance norm (Ulyanov et al., 2016), accumulate gradients over 4 forward passes before each parameter update, and switch from float32 to bfloat16 to further reduce memory. These changes maintain training dynamics comparable to the 1◦ and 1/2◦ configurations. Training uses PyTorchDistributedDataParallelacross 8 A100 GPUs; Table 2 summarizes hardware ...

    2016

  29. [29]

    B.2 Spectral Analysis The isotropic temperature and EKE spatial spectra are presented in the main text (Fig- ure 9)

    show a similar resolution dependence, with progressively finer spatial detail emerging at higher resolution. B.2 Spectral Analysis The isotropic temperature and EKE spatial spectra are presented in the main text (Fig- ure 9). Here we provide additional temporal spectra and autocorrelation diagnostics. The 28 Samudra 2 temporal EKE and temperature power sp...

    work page arXiv 2000