Pith. sign in

REVIEW 2 major objections 5 minor 54 references

Kastor: An efficient fine-tuning strategy for generative emulation of PDE simulations

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A fine-tuning recipe turns a deterministic physics model into a generative surrogate, cutting forecasting error by 42.9 percent on average across ten 2D PDE simulation datasets.

desk verdict Kastor is a solid empirical contribution to generative PDE surrogate training, but the headline 42.9% fCRPS win is anchored to a self-reproduced baseline and a τ=4-subsampled evaluation that never scores the upsampler's intermediate frames. read the letter →

arxiv 2608.06107 v1 pith:MFK6II73 submitted 2026-08-06 cs.LG cs.NAmath.NA

Guillaume Couairon , Alexis Jacq , Yu-Han Wu , Renu Singh , Yana Hasson , Quentin Berthet , Romuald Elie This is my paper · ORCID
classification cs.LGcs.NAmath.NA
keywords PDEsurrogatesgenerativeemulationfoundationmodelfine-tuningensembleforecastingCRPSmeanpredictionregularizationspectralconsistencytemporalsuper-resolution
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 claims that a pretrained deterministic physics model, used as is for point forecasts, can be repurposed into a generative ensemble emulator that is more accurate, better calibrated, and cheaper to roll out than the standard recipe of fine-tuning the same backbone. The authors propose three ingredients: rolling out the causal model with a time stride and then filling in skipped frames with a small non-causal upsampler; a regularization that forces the generative model, when noise is zero, to reproduce the conditional mean of the next state; and a loss that matches spatial gradients of the forecast to the ground truth. They report that the resulting model reduces forecasting error by 42.9 percent on average across ten 2D PDE simulation datasets, improves spread-skill calibration on most of those datasets, and lowers inference cost relative to the fine-tuning baseline. A sympathetic reader would care because generative surrogates are wanted for risk assessment and uncertainty quantification, and this is a recipe for obtaining them from existing deterministic foundation models rather than training large generative models from scratch.

What carries the argument

The argument runs on three mechanisms. First, a two-stage inference scheme: a causal transformer predicts states every $\tau$ frames ($\tau = 4$), and a half-size non-causal transformer fills the missing intermediate states using surrounding context, which cuts error accumulation and reduces rollout cost by roughly 37.5 percent at long horizons. Second, Mean Prediction Regularization (MPR): the generative model is trained so that conditioning on null noise $\epsilon = 0$ yields the conditional mean of the next-state distribution, supervised by an $\ell^1$ loss to the ground-truth state difference; this stabilizes both FGN and diffusion variants and is what makes FGN competitive. Third, the Gradient Difference Loss (GDL), which adds a loss on spatial derivatives of the predicted field, improving accuracy and bringing the radially-averaged power spectrum closer to the ground truth. These are implemented on the Walrus backbone with AdaLN noise injection, patch jittering as a second stochasticity source, and a CRPS scoring rule for the generative objective.

What would settle it

Re-run the same evaluation protocol (eight ensemble members, 32-step rollouts, two starting contexts per trajectory) on the official Walrus checkpoint; if the official baseline matches or beats the reproduction on Acoustic Scattering Discontinuity, Acoustic Scattering Maze, and PlanetSWE, the averaged 42.9 percent fCRPS reduction will not hold.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the failure modes of generative PDE emulators are largely training and inference choices, not an inherent limit of the backbone. Casting the pretrained Walrus architecture as a single-pass generative model trained with the CRPS scoring rule works only when the model is also constrained to predict the deterministic mean at zero noise; without this Mean Prediction Regularization, the FGN variant is worse than a simple mean-absolute-error baseline. With that constraint, plus spatial gradient matching and a strided rollout with temporal super-resolution, the authors obtain a model that beats their retrained Walrus fine-tuning reference on forecasting error, ensemble calibration, and spectral error, and does so with fewer autoregressive steps. The 42.9 percent average fCRPS reduction and the 8-of-10 datasets VRMSE win are the paper's headline numbers.

Load-bearing premise

The headline 42.9 percent reduction is computed against the authors' retrained version of the Walrus fine-tuning baseline, not against the official published Walrus checkpoint, and that reproduction performs noticeably worse on three datasets.

Editorial extensions

If this is right

  • Pretrained deterministic physics models can be converted into generative ensemble emulators without architectural changes, so the recipe transfers to any backbone with attention layers.
  • Strided rollouts with temporal super-resolution reduce both error accumulation and inference cost, making long-horizon ensembles practical.
  • MPR is a general regularizer: it improves both FGN and diffusion models, and likely applies to other generative families trained with scoring rules or denoising objectives.
  • Gradient matching restores high-frequency content, so models trained with it produce power spectra closer to ground truth than their unregularized counterparts.
  • The improved calibration and spectral fidelity make the emulator usable for uncertainty quantification and for computing time-integrated statistics over full trajectories.

Reading between the lines

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

  • The success of MPR suggests that in PDE emulation the generative component should be anchored to the deterministic mode; extending the null-noise constraint to flow-matching or latent diffusion objectives is a direct test.
  • Because the time-stride optimum varies by dataset and lead time, a rollout policy that starts with $\tau=1$ for the first few steps and switches to $\tau=4$ later could beat either fixed stride; the paper notes this combination as future work.
  • The headline 42.9 percent gain is relative to the authors' reproduction of the Walrus fine-tuning baseline; if the official checkpoint is used instead, the gain on acoustic scattering and PlanetSWE is likely to be smaller, though the directional improvements from MPR and GDL should persist.
  • GDL on spatial gradients could be extended to divergence or curl matching for incompressible flows, where those quantities are conserved and would directly penalize unphysical artifacts.
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.

Referee Report

2 major / 5 minor

Summary. The paper proposes Kastor, a recipe for fine-tuning the Walrus physics foundation model into a generative ensemble emulator for PDE simulation data. Three components are introduced: a two-stage inference scheme that first rolls out a causal model with a large time stride and then fills in intermediate states with a smaller non-causal upsampler; Mean Prediction Regularization (MPR), which adds a regression loss on the network output at zero noise conditioning; and a Gradient Difference Loss (GDL) that matches spatial gradients of forecasts. The authors report a 42.9% average reduction in fCRPS relative to their re-trained Walrus fine-tuning baseline across ten 2D datasets from The Well, along with improved ensemble calibration and spectral fidelity at lower inference cost, and they show through ablations that MPR and GDL improve both FGN and diffusion emulators.

Significance. Strengths: the paper includes extensive per-dataset and per-lead-time ablation figures, a transparent cost model in Appendix B.2, and an honest limitations section. The MPR and GDL ablations in Figures 7 and 9 provide internal support for the proposed components, and the use of the public The Well benchmark makes the empirical claims reproducible in principle. If the headline comparison were made against the official Walrus checkpoint and dense trajectories, the claimed 42.9% improvement would be a practically significant result. At present, however, the headline number is tied to a self-retrained baseline and to sub-sampled lead times, so the external significance is not yet fully established.

major comments (2)
  1. [Section 3.6, Tables 1-2] The headline 42.9% fCRPS reduction is computed on trajectories sub-sampled at tau=4, i.e., only lead times 4,8,...,32. This excludes lead times 1-3, where Section 4.1 and Figure 23 show the tau=1 baseline is often better, and it never scores the interpolated frames produced by the non-causal upsampler that is central to the proposed two-stage inference scheme. Consequently, the abstract's claim that Kastor 'reduces forecasting error' is not yet established for complete trajectories. Please report the skill scores on dense full-length trajectories (all lead times 1..32, including the upsampler outputs) and state explicitly whether the 42.9% figure refers only to sub-sampled lead times.
  2. [Appendix C.1, Figure 21] The reference model in the skill score is the authors' reproduction of the Walrus fine-tuning methodology, not an official Walrus checkpoint. Appendix C.1 concedes that the reproduction performs noticeably worse than reported for Acoustic Scattering Discontinuity, Acoustic Scattering Maze, and PlanetSWE; for Acoustic Scattering Maze, Kastor-L (VRMSE 0.038) lies between the reproduction (0.047) and the reported Walrus (0.035), and for PlanetSWE the reproduction gap is large. Since the headline 42.9% fCRPS reduction and the 8/10 VRMSE claim are both measured against this reproduction, the advantage over the published Walrus model may be materially smaller on those datasets. Please evaluate against the official Walrus checkpoint or, if it is unavailable, provide a second analysis using the reported Walrus numbers as a reference.
minor comments (5)
  1. [Section 3.6] The sentence 'For consistency across datasets and experiments, we always report aggregated metrics on trajectories sub-sampled at tau=4' should be moved into the abstract or headline claim so readers are not surprised when Tables 1-2 list only tau=4,8,...
  2. [Section 3.2, Eq. (2)] MPR is presented as a novel regularization, but it is essentially a consistency constraint between the generative output at null noise and a regression target; please discuss its relation to consistency training and to the residual modeling baselines in Appendix C.8.
  3. [Figure 4 caption] The per-lead-time curves for tau>1 use different evaluation protocols (sub-sampled versus upsampled), and the figure caption should state clearly how lead times 1-3 are obtained for those models.
  4. [Appendix B.2] The cost comparison assumes KV caching for the causal baseline and a half-size upsampler; please include a sensitivity analysis around the upsampler size and the caching assumptions.
  5. [Abstract and Table 2] The abstract says 'outperforms Walrus for 8 out of 10 datasets on VRMSE', but Table 2 reports Ensemble Mean RMSE, not VRMSE; clarify the metric used for the 8/10 claim.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular reduction found; fCRPS headline is empirical but anchored to a self-retrained Walrus baseline and sub-sampled trajectories.

full rationale

We walked the claimed derivation chain. The two-stage inference scheme (Sec. 3.1) trains a strided causal model and a non-causal upsampler against ground-truth states with standard losses; it contains no step where the predicted quantity is defined in terms of the evaluation metric. MPR (Eqs. 1-2) is an L1 regression loss on the null-noise output, supervised directly by the ground-truth state difference; GDL (Eq. 5) is a spatial-gradient matching loss; and the FGN/CRPS objective (Eqs. 3-4) is a strictly proper scoring rule trained on ground-truth references. None of these reduce, by construction, to the reported fCRPS or VRMSE improvements; the gains are empirical outcomes of training, not identities. The main caveat is benchmark anchoring, not circularity: the 42.9% fCRPS reduction is computed against the authors' retrained Walrus fine-tuning baseline (Sec. 3.7), and Appendix C.1 acknowledges that this reproduction performs "noticeably worse" than the reported Walrus on Acoustic Scattering Discontinuity, Acoustic Scattering Maze, and PlanetSWE. If the official checkpoint were used, the headline gap would likely shrink. This is a baseline-fairness/external-validity concern, not a by-construction equivalence. Similarly, Sec. 3.6 states that aggregated metrics are always reported on trajectories sub-sampled at tau=4, and Tables 1-2 list only lead times 4,8,...,32; this may overstate full-trajectory benefit and skips early lead times where the paper admits tau=1 is better. Again, this is a reporting/evaluation gap rather than a circular reduction. Self-citations appear in Appendix C.8 (residual models, Couairon et al. 2026) and Appendix B.4 (distillation, Jacq et al. 2026), but they are minor baselines or forward-looking mentions, not load-bearing premises. No uniqueness theorem is imported from the authors' prior work, no ansatz is smuggled in via citation, and no known result is merely renamed. We therefore find no significant circularity; the score of 2 reflects the minor self-citations and the externally anchored baseline caveat, not a circular derivation.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the pretrained Walrus backbone, the generative modeling assumptions of FGN and diffusion, hand-chosen loss weights and per-dataset time strides, and the Well benchmark data. No new physical entities, forces, or conserved quantities are introduced. The main load-bearing choices are empirical and validation-selected rather than derived.

free parameters (5)
  • per-dataset time stride tau = tau=4 for 8 of 10 datasets; tau=1 for Active Matter and Rayleigh-Benard
    Chosen on validation from the ablation in Section 4.1 and Figures 4 and 23; directly affects error accumulation and all reported rollout metrics.
  • MPR loss weight lambda_MPR = 100
    Set by hand in the combined loss in Eq. 14; supervises the zero-noise pass against ground truth; no sensitivity analysis is provided.
  • CRPS loss weight lambda_CRPS = 100
    Set by hand in Eq. 14; balances the generative scoring rule against MPR and GDL.
  • GDL loss weight lambda_GDL = 2000
    Chosen because spatial gradient magnitudes are small, stated in Section 3.8; no sensitivity analysis is provided.
  • diffusion baseline step count = 5
    Selected after experiments showing degradation beyond 5 steps in Figure 31; affects the FGN-versus-diffusion comparison.
assumptions (5)
  • domain assumption Walrus pretrained weights provide a useful prior for all 10 The Well datasets used.
    All fine-tuning starts from Walrus weights as described in Sections 3 and 3.8; if this prior fails, the method's claimed transferability breaks.
  • domain assumption The conditional distribution of the next PDE state can be represented by injecting a 32-dimensional Gaussian noise via AdaLN plus patch jittering.
    Core FGN construction in Section 3.3; no proof that this low-dimensional latent family covers the target stochasticity.
  • standard math CRPS computed against a single ground-truth sample with the fair correction is a valid training signal.
    Scoring-rule theory from Gneiting and Raftery 2007 and Ferro 2014, used in Section 3.3 and Eq. 4.
  • domain assumption Global mean and standard deviation normalization is the right fine-tuning choice despite the pretrained model using sample-wise normalization.
    Tested in Appendix C.2 and chosen as default; it changes the input distribution the pretrained weights were optimized for.
  • domain assumption The 10 selected Well datasets are representative of the target PDE emulation use cases.
    All empirical claims are limited to these datasets; Section B.1 excludes three 2D datasets for stated practical reasons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Kastor: An efficient fine-tuning strategy for generative emulation of PDE simulations." pith.science (2026). https://pith.science/paper/MFK6II73

@misc{pith2026260806107,
  author       = {Pith},
  title        = {Pith review of: Kastor: An efficient fine-tuning strategy for generative emulation of PDE simulations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MFK6II73}},
  note         = {Machine review of arXiv:2608.06107}
}
read the original abstract

Machine learning offers a promising avenue to accelerate physical simulations by replacing computationally expensive traditional Partial Differential Equation (PDE) solvers with fast, differentiable surrogate models. However, standard auto-regressive ML emulators often suffer from error accumulation over long horizons and struggle to capture the stochasticity of complex physical systems. In this paper, we propose Kastor, a comprehensive methodology to adapt a deterministic physics foundation model into a highly efficient and accurate generative surrogate. First, we introduce a two-stage inference scheme that combines a large-stride causal auto-regressive model with a non-causal temporal super-resolution network, significantly reducing error accumulation while minimizing computational cost. Second, we present Mean prediction regularization (MPR), a novel training objective that constrains the generative model to predict the deterministic distribution mean under null noise conditioning. This regularization dramatically improves the performance and stability of both Functional Generative Networks (FGN) and diffusion-based emulators. Finally, we demonstrate that incorporating spatial gradient matching improves the accuracy and physical fidelity of the simulations as measured by power spectrum density. Extensive evaluations on diverse simulation datasets of the benchmark The Well show that with these components, our model outperforms competing methods in forecasting accuracy, spectral consistency, and computational efficiency. Our model achieves a 42.9% average reduction in forecasting compared to our reference based on the Walrus finetuning methodology, and outperforms Walrus for 8 out of 10 datasets on variance-normalized RMSE (VRMSE).

Figures

Figures reproduced from arXiv: 2608.06107 by the authors.

Figure 1
Figure 1. Average relative reduction of forecasting error [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Diagram of our proposed method to recover full trajectories with non-causal temporal super-resolution (middle), [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Diagram of Mean Prediction Regularization principle (left) and its implementation with FGN (right). The training [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (28 more)
Figure 4
Figure 4. Figure 4: Ensemble Mean RMSE skill score ↑ per lead time, for four representative datasets and MAE-trained models (size S) at strides τ ∈ {1, 2, 4, 8}. While stride T = 1 often performs best at early lead times, on a majority of datasets using larger strides is beneficial for lo…
Figure 6
Figure 6. Figure 6: Qualitative comparison between different time [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Performance of S models trained with and without [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Comparison of FGN models trained with different [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 10
Figure 10. Figure 10: Qualitative comparison of Kastor-M (20 input layers) and Kastor-L (40 input layers) on samples from the Periodic [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparison of Kastor-M (20 input layers) and Kastor-L (40 input layers) on the center crop of a validation [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison of Kastor-M (20 input layers) and Kastor-L (40 input layers) on samples from the [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Additional qualitative examples comparing different time strides on validation samples from the Euler Multi-quadrants [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Qualitative examples comparing different time strides on two validation samples from the Rayleigh-Bénard dataset. [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: Qualitative examples comparing different time strides on two validation samples from the [PITH_FULL_IMAGE:figures/full_fig_p016_15.png]
Figure 16
Figure 16. Figure 16: Qualitative comparison of forecasting on the [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]
Figure 17
Figure 17. Figure 17: Qualitative comparison of FGN regularization on two samples from the [PITH_FULL_IMAGE:figures/full_fig_p018_17.png]
Figure 18
Figure 18. Figure 18: Qualitative study of the effect of regularization of FGN models on the center-crops of three samples from the [PITH_FULL_IMAGE:figures/full_fig_p019_18.png]
Figure 19
Figure 19. Figure 19: Qualitative comparison of FGN regularization methods on samples from the Acoustic Scattering Maze dataset across [PITH_FULL_IMAGE:figures/full_fig_p020_19.png]
Figure 20
Figure 20. Figure 20: Diagram of Mean Prediction Regularization applied to diffusion modeling. [PITH_FULL_IMAGE:figures/full_fig_p023_20.png]
Figure 21
Figure 21. Figure 21: Median VRMSE for lead times T = 1 − 20 for our reproduction of the Walrus finetuning methodology, which we used as reference for skill score computations, Kastor-L w/ upsampling compared to the scores reported in (McCabe et al. 2025). Our reproduction performs similar…
Figure 22
Figure 22. Figure 22: RMSE relative improvement (skill score) of the Ensemble mean for various configurations. Patch jittering, global [PITH_FULL_IMAGE:figures/full_fig_p024_22.png]
Figure 23
Figure 23. Figure 23: Per-dataset comparison of models trained with different time stride [PITH_FULL_IMAGE:figures/full_fig_p026_23.png]
Figure 24
Figure 24. Figure 24: Per-dataset comparison of a τ = 1 MAE-S model trained with supervision on 4 consecutive rollout steps, compared to MAE-S models with time strides τ = {1, 4}. Models are evaluated with Ensemble Mean RMSE skill score per lead time. Multi-step finetuning improves perform…
Figure 25
Figure 25. Figure 25: Improvement in forecasting error (EnsMeanRMSE skill score) for our S (20 layers, 2 input frames), M (20 layers, 4 [PITH_FULL_IMAGE:figures/full_fig_p026_25.png]
Figure 26
Figure 26. Figure 26: Forecasting error improvement (fCRPS skill [PITH_FULL_IMAGE:figures/full_fig_p027_26.png]
Figure 29
Figure 29. Figure 29: While residual modeling brings improvements for [PITH_FULL_IMAGE:figures/full_fig_p027_29.png]
Figure 27
Figure 27. Figure 27: Overall improvement (higher) of forecasting error (fCRPSss and EnsMeanRMSEss), spectral consistency (LSDss), [PITH_FULL_IMAGE:figures/full_fig_p028_27.png]
Figure 28
Figure 28. Figure 28: Radially-averaged power spectra (RAPSD) for [PITH_FULL_IMAGE:figures/full_fig_p029_28.png]
Figure 30
Figure 30. Figure 30: fCRPSsf scores per lead time for 2 and 4 input [PITH_FULL_IMAGE:figures/full_fig_p029_30.png]
Figure 29
Figure 29. Figure 29: Forecasting error (fCRPSss) of our models com [PITH_FULL_IMAGE:figures/full_fig_p029_29.png]
Figure 31
Figure 31. Figure 31: Impact of the number of diffusion steps on fCRPSss (top) and LSD (bottom). The quality of rollouts increases with [PITH_FULL_IMAGE:figures/full_fig_p030_31.png]
Figure 32
Figure 32. Figure 32: Impact of the number of diffusion steps, when using GDL. With GDL, the quality of rollouts increases with the [PITH_FULL_IMAGE:figures/full_fig_p030_32.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 27 canonical work pages

  1. [1]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  2. [2]

    2025 , journal=

    Walrus: A Cross-Domain Foundation Model for Continuum Dynamics , author=. 2025 , journal=

    work page 2025

  3. [3]

    Ohana, Ruben and McCabe, Michael and Meyer, Lucas and Morel, Rudy and Agocs, Fruzsina J and Beneitez, Miguel and Berger, Marsha and Burkhart, Blakesley and Dalziel, Stuart B and Fielding, Drummond B and others , booktitle=. The

    work page

  4. [4]

    International Conference on Learning Representations , editor =

    Overtone: Cyclic Patch Modulation for Clean, Efficient, and Flexible Physics Emulators , author=. International Conference on Learning Representations , editor =

    work page

  5. [5]

    arXiv preprint arXiv:2603.05598 , year=

    On the Value of Tokeniser Pretraining in Physics Foundation Models , author=. arXiv preprint arXiv:2603.05598 , year=

    work page arXiv

  6. [6]

    arXiv preprint arXiv:2602.00884 , year=

    Test-time Generalization for Physics through Neural Operator Splitting , author=. arXiv preprint arXiv:2602.00884 , year=

    arXiv

  7. [7]

    2025 , organization=

    Lightweight Fourier Neural Operator for Time-Dependent Partial Differential Equations , author=. 2025 , organization=

    work page 2025

  8. [8]

    Advances in Neural Information Processing Systems , editor=

    Mixture-of-experts operator transformer for large-scale pde pre-training , author=. Advances in Neural Information Processing Systems , editor=

    work page

Show all 54 references
  1. [9]

    arXiv preprint arXiv:2509.21670 , year=

    MORPH: PDE foundation models with arbitrary data modality , author=. arXiv preprint arXiv:2509.21670 , year=

  2. [10]

    JMLR , volume =

    Nikola Kovachki and Zongyi Li and Burigede Liu and Kamyar Azizzadenesheli and Kaushik Bhattacharya and Andrew Stuart and Anima Anandkumar , title =. JMLR , volume =

  3. [11]

    Advances in Neural Information Processing Systems , editor=

    Multiple Physics Pretraining for Spatiotemporal Surrogate Models , author=. Advances in Neural Information Processing Systems , editor=. 2024 , volume=

    2024

  4. [12]

    Forty-third International Conference on Machine Learning , year=

    Probabilistic Retrofitting of Learned Simulators , author=. Forty-third International Conference on Machine Learning , year=

  5. [13]

    Advances in Neural Information Processing Systems , editor=

    ENMA: Tokenwise Autoregression for Continuous Neural PDE Operators , author=. Advances in Neural Information Processing Systems , editor=

  6. [14]

    arXiv preprint arXiv:2606.30414 , year=

    Diffusion Fine-tuning with Rewarded Moment Matching Distillation , author=. arXiv preprint arXiv:2606.30414 , year=

    arXiv

  7. [15]

    arXiv preprint arXiv:2506.17774 , year=

    Physix: A foundation model for physics simulations , author=. arXiv preprint arXiv:2506.17774 , year=

    arXiv

  8. [16]

    Ye, Zhanhong and Liu, Zining and Wu, Bingyang and Jiang, Hongjie and Chen, Leheng and Zhang, Minyan and Huang, Xiang and Zou, Qinghe Meng and Liu, Hongsheng and Dong, Bin and others , journal=

  9. [17]

    arXiv preprint arXiv:2511.21861 , year=

    Towards a Foundation Model for Partial Differential Equations Across Physics Domains , author=. arXiv preprint arXiv:2511.21861 , year=

  10. [18]

    arXiv preprint arXiv:2509.13805 , year=

    Towards a physics foundation model , author=. arXiv preprint arXiv:2509.13805 , year=

    arXiv

  11. [19]

    arXiv preprint arXiv:2604.09041 , year=

    U-Cast: A Surprisingly Simple and Efficient Frontier Probabilistic AI Weather Forecaster , author=. arXiv preprint arXiv:2604.09041 , year=

    arXiv

  12. [20]

    Advances in Neural Information Processing Systems , editor=

    Lost in latent space: An empirical study of latent diffusion models for physics emulation , author=. Advances in Neural Information Processing Systems , editor=

  13. [21]

    arXiv preprint arXiv:2506.10772 , year=

    Skillful joint probabilistic weather forecasting from marginals , author=. arXiv preprint arXiv:2506.10772 , year=

    arXiv

  14. [22]

    Quarterly Journal of the Royal Meteorological Society , volume=

    Fair scores for ensemble forecasts , author=. Quarterly Journal of the Royal Meteorological Society , volume=. 2014 , publisher=

    2014

  15. [23]

    Journal of the American statistical Association , volume=

    Strictly proper scoring rules, prediction, and estimation , author=. Journal of the American statistical Association , volume=. 2007 , publisher=

    2007

  16. [24]

    Perez, Ethan and Strub, Florian and De Vries, Harm and Dumoulin, Vincent and Courville, Aaron , booktitle=. Fi

  17. [25]

    The Fourth International Conference on Learning Representations , year=

    Deep multi-scale video prediction beyond mean square error , author=. The Fourth International Conference on Learning Representations , year=

  18. [26]

    BCAT: A block causal transformer for

    Liu, Yuxuan and Sun, Jingmin and Schaeffer, Hayden , journal=. BCAT: A block causal transformer for

  19. [27]

    2025 , editor=

    Morel, Rudy and Han, Jiequn and Oyallon, Edouard , booktitle=. 2025 , editor=

    2025

  20. [28]

    2026 , publisher=

    Foundation neural operators: A survey on pretraining methods, the data ecosystem, and efficient adaptation , author=. 2026 , publisher=

    2026

  21. [29]

    arXiv preprint arXiv:2511.20798 , year=

    Physics Steering: Causal Control of Cross-Domain Concepts in a Physics Foundation Model , author=. arXiv preprint arXiv:2511.20798 , year=

  22. [30]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    NESTOR: A Nested MOE-based Neural Operator for Large-Scale PDE Pre-Training , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  23. [31]

    Liu, Yuxuan and Sun, Jingmin and He, Xinjie and Pinney, Griffin and Zhang, Zecheng and Schaeffer, Hayden , journal=

  24. [32]

    arXiv preprint arXiv:2512.23056 , year=

    PI-MFM: Physics-informed multimodal foundation model for solving partial differential equations , author=. arXiv preprint arXiv:2512.23056 , year=

  25. [33]

    arXiv preprint arXiv:2606.11657 , year=

    Sparse probes and murky physics: a case study of interpretability challenges in a foundation model for continuum dynamics , author=. arXiv preprint arXiv:2606.11657 , year=

    arXiv

  26. [34]

    Science Advances , volume =

    Guillaume Couairon and Renu Singh and Anastase Charantonis and Christian Lessig and Claire Monteleoni , title =. Science Advances , volume =. 2026 , doi =

    2026

  27. [35]

    2024 , url =

    Keller Jordan and Yuchen Jin and Vlado Boza and Jiacheng You and Franz Cesista and Laker Newhouse and Jeremy Bernstein , title =. 2024 , url =

    2024

  28. [36]

    arXiv preprint arXiv:1412.6980 , year=

    Adam: A method for stochastic optimization , author=. arXiv preprint arXiv:1412.6980 , year=

    arXiv

  29. [37]

    Geoscientific Model Development , volume=

    An evolving Coupled Model Intercomparison Project phase 7 (CMIP7) and Fast Track in support of future climate assessment , author=. Geoscientific Model Development , volume=. 2025 , publisher=

    2025

  30. [38]

    Nature Reviews Physics , volume=

    Accelerating fusion research via supercomputing , author=. Nature Reviews Physics , volume=. 2025 , publisher=

    2025

  31. [39]

    2005 , publisher=

    Numerical solution of partial differential equations: an introduction , author=. 2005 , publisher=

    2005

  32. [40]

    Science , volume=

    Learning skillful medium-range global weather forecasting , author=. Science , volume=. 2023 , publisher=

    2023

  33. [41]

    International Journal of Hydromechatronics , volume=

    Surrogate models in machine learning for computational stochastic multi-scale modelling in composite materials design , author=. International Journal of Hydromechatronics , volume=. 2022 , publisher=

    2022

  34. [42]

    Renewable Energy , volume=

    Data-driven modal parameterization for robust aerodynamic shape optimization of wind turbine blades , author=. Renewable Energy , volume=. 2024 , publisher=

    2024

  35. [43]

    Advances in Neural Information Processing Systems , volume=

    Sobolev training for neural networks , author=. Advances in Neural Information Processing Systems , volume=

  36. [44]

    arXiv preprint arXiv:2511.20455 , year=

    Fluid Intelligence: A Forward Look on AI Foundation Models in Computational Fluid Dynamics , author=. arXiv preprint arXiv:2511.20455 , year=

  37. [45]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  38. [46]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  39. [47]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  40. [48]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  41. [49]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    1984 doi

  42. [50]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  43. [51]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  44. [52]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  45. [53]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  46. [54]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.