Pith. sign in

REVIEW 4 major objections 5 minor 10 cited by

ArchesWeather & ArchesWeatherGen: a deterministic and generative model for efficient ML weather forecasting

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A two-stage flow-matching weather model beats IFS ENS and NeuralGCM on most headline variables while training for about 9 V100-days.

desk verdict Solid and reproducible generative weather forecasting paper, but the abstract overclaims and should be corrected before publication. read the letter →

arxiv 2412.12971 v1 pith:W2VGJJXL submitted 2024-12-17 cs.LG

Guillaume Couairon , Renu Singh , Anastase Charantonis , Christian Lessig , Claire Monteleoni This is my paper · ORCID
classification cs.LG
keywords weatherforecastingprobabilisticflowmatchingdiffusionmodelsensembleERA5transformerresidualmodeling
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 claims that a probabilistic weather model can be built cheaply by training a deterministic forecaster first and then a flow-matching generative model on the residuals. The deterministic model, ArchesWeather, is a transformer whose Cross-Level Attention lets all pressure levels interact; the generative model, ArchesWeatherGen, learns to turn Gaussian noise into realistic residual weather states that are added to the deterministic prediction. The paper argues that this makes ArchesWeatherGen a true stochastic emulator of ERA5: its members have realistic small-scale structure, its rank histograms are nearly flat, and its spread-skill ratio is close to 1. If the claim holds, credible ensemble weather forecasting no longer requires high resolution or massive computational budgets.

What carries the argument

The load-bearing object is the residual flow-matching transition model. A deterministic model $f_\theta$ trained with MSE loss approximates the conditional mean $\mathbb{E}[x_{t+\delta}\mid x_t]$; residuals $r_{t+\delta}=(x_{t+\delta}-f_\theta(x_t))/\sigma$ are renormalized to unit variance, and a denoiser $g_\theta$ is trained with the flow-matching loss on $(1-s)r+s\epsilon$ for $s\in[0,1]$. Sampling follows the neural ODE $dz_s=(g_\theta(z_s)-z_s)ds$ with 25 Euler steps. Two corrections carry the calibration: OOD fine-tuning on 2019 data that the deterministic model never saw, and input noise scaling by a factor of 1.05 to compensate for deterministic-model overfitting. The architectural enabler is Cross-Level Attention (CLA), column-wise attention along the vertical dimension combined with horizontal 2D windows, which gives a full vertical receptive field at $O(d^2)$ parameters instead of $O(Z^2d^2)$.

What would settle it

Take ArchesWeatherGen with the global noise scaling fixed at 1.05 and compute its fair CRPS and spread-skill ratio on a test year outside the tuning loop, for example 2021; if the spread-skill ratio moves substantially away from 1 or CRPS degrades relative to IFS ENS, the claim that it is a true stochastic emulator of ERA5 would fail.

Watch

Extended reading notes

Core claim

The paper's central claim is that a probabilistic weather model can match or beat operational ensemble systems by decomposing the forecast distribution into a deterministic mean and a generative residual. The deterministic part, ArchesWeather, is a 1.5-degree Swin U-Net transformer whose Cross-Level Attention layer gives a global receptive field along the vertical axis; a 4-member ensemble of these models approximates the conditional mean $\mathbb{E}[x_{t+\delta}\mid x_t]$. The generative part, ArchesWeatherGen, is a flow-matching denoiser trained on renormalized residuals $r_{t+\delta}=(x_{t+\delta}-f_\theta(x_t))/\sigma$. At inference, the denoiser maps scaled Gaussian noise to a residual sample, adds it to the deterministic prediction, and repeats autoregressively. The paper argues that this design makes ArchesWeatherGen a true stochastic emulator of ERA5, with power spectra and activity close to the reanalysis, near-flat rank histograms, and better fair CRPS, energy score, Brier score, and ensemble-mean RMSE than IFS ENS and NeuralGCM on all WeatherBench headline variables except geopotential.

Load-bearing premise

The claim stands or falls on the assumption that after OOD fine-tuning on 2019 data and a noise scaling factor of 1.05, the residuals the model sees at test time come from the same distribution as the residuals it was trained on; any shift in the deterministic model's overfitting rate or in the residual distribution breaks the spread-skill calibration.

Editorial extensions

If this is right

  • A 4-member deterministic ensemble at 1.5 degrees can match or beat much larger ensembles on ensemble-mean RMSE, so credible probabilistic forecasts become accessible with academic compute.
  • Sampling from ArchesWeatherGen restores small-scale variability that deterministic forecasts smooth out, with power spectra and activity levels close to ERA5.
  • OOD fine-tuning and noise scaling lift the spread-skill ratio from about 0.85 to 0.96-0.98, which is what makes the 'true stochastic emulator' claim hold.
  • Flow matching outperforms the DDPM variant by roughly 4% in relative CRPS, suggesting that two-stage residual training with flow matching is a better recipe than one-stage diffusion.
  • Since the method surpasses IFS ENS and NeuralGCM on most variables but not geopotential, hybrid dynamical-core models may retain an edge on smooth fields like Z500.

Reading between the lines

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

  • The same residual-flow-matching recipe could upgrade any existing MSE-trained deterministic weather model to a probabilistic ensemble model, since the deterministic model only needs to provide a mean estimate.
  • A per-variable noise scaling coefficient could remove the remaining slight underdispersion in variables such as T2m, a direct testable modification of the single global 1.05 scaling.
  • If the two-stage decomposition holds at higher resolution, it could bring the same cost savings to km-scale stochastic emulation and data assimilation, where full generative training is currently expensive.
  • The 'true stochastic emulator' claim is about matching ERA5 at 1.5 degrees, not about physical correctness of individual members; spectral realism and flat rank histograms do not by themselves guarantee that member dynamics obey atmospheric equations.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes ArchesWeather, a 1.5-degree-resolution transformer-based deterministic weather model, and ArchesWeatherGen, a flow-matching generative model trained on residuals from ArchesWeather to produce probabilistic forecasts. The deterministic model introduces a cross-level attention (CLA) mechanism to replace Pangu-Weather's local 3D attention and achieves competitive RMSE with a small training budget. The generative model is trained in two stages: first on 1979-2018 residuals, then fine-tuned on 2019 residuals, with a manually chosen noise-scaling coefficient of 1.05 to correct underdispersion. The authors report that ArchesWeatherGen outperforms IFS ENS and NeuralGCM on most WeatherBench variables in fCRPS, Ensemble Mean RMSE, and Brier score, with much lower compute cost, and they commit to releasing code and evaluation pipelines.

Significance. The contribution is potentially significant: the two-stage residual flow-matching approach is computationally efficient, the ablation study clearly isolates the effects of OOD fine-tuning and noise scaling, and the paper is unusually transparent about residual underdispersion and about the heuristic choice of the noise-scaling coefficient. If the reported scores are reproducible, the work provides a low-cost probabilistic forecasting baseline that is competitive with much more expensive systems. However, the headline claim of superiority over IFS ENS on all WeatherBench headline variables is contradicted by the paper's own per-variable fCRPS results for T2m at lead times beyond 9 days, and the 'true stochastic emulator' claim relies on a manually calibrated scalar rather than on a validation-based derivation. The core methodology is sound enough to warrant a major revision rather than rejection.

major comments (4)
  1. [Abstract; Section 4.3, Figure 10] The Abstract claims that ArchesWeatherGen 'surpasses IFS ENS and NeuralGCM on all WeatherBench headline variables (except for NeuralGCM's geopotential)'. Section 4.3 and Figure 10 report that ArchesWeatherGen is 'slightly worse only for T2m at lead times longer than 9 days' relative to IFS ENS. Since T2m is one of the headline variables, the abstract's 'all ... variables' statement is internally inconsistent with the paper's own per-variable results. Please revise the Abstract (and the matching claim in Section 1) to include this per-variable caveat, or provide a summary metric over all nine headline variables that supports the global statement.
  2. [Section 3.3, 'Protocol for training ArchesWeatherGen'; Section 3, evaluation protocol] The generative model is fine-tuned on 2019 data, which is the year immediately preceding the 2020 test year. As a result, the 2020 evaluation is not fully out-of-sample for ArchesWeatherGen: the OOD fine-tuning phase is in effect calibration on a temporally adjacent year. The ablation in Section 4.4 (Figure 12) shows that OOD fine-tuning improves all ensemble metrics, so the reported gains over IFS ENS may be partly attributable to this adjacent-year tuning. Please add an evaluation on a later year (e.g., 2021) or otherwise demonstrate that the 2019 fine-tuning does not inflate the 2020 scores; this is necessary to support the claim that the model is a general stochastic emulator of ERA5 rather than a model calibrated to the immediate pre-test year.
  3. [Section 3.3; Section 4.4, Figures 12 and 16] The spread-skill ratio, which is the basis for the 'true stochastic emulator' claim (Section 3.4), is calibrated with a manually chosen noise-scaling coefficient rho = 1.05, stated in Section 3.3 to 'roughly correspond to the percentage of overfitting observed'. Figure 16 shows that per-variable spread-skill ratios still deviate substantially from 1 (e.g., T2m and Z500 at several lead times), and the coefficient is a single scalar applied uniformly to all variables and lead times. The paper should either derive rho from an independent validation criterion, report the sensitivity of the headline comparisons to rho, or soften the 'true stochastic emulator' wording to reflect that the dispersion is calibrated rather than learned.
  4. [Section 4.3, 'Summary of metrics'; Figure 9] The summary skill-score plots in Figure 9 average over only the five upper-air headline variables (Z500, Q700, T850, U850, V850), as the text explicitly states, while the Abstract claims superiority on 'all WeatherBench headline variables'. Because the headline set also includes T2m, SP, U10m, and V10m, the averaged upper-air plot does not by itself support the broad claim, and Figure 10 shows that the per-variable picture is more mixed. Please either report the summary metric over the full nine-variable headline set or restrict the global claim to the upper-air variables.
minor comments (5)
  1. [Section 3.3, Eq. (6)] Equation (6) conditions the generative model on xt and f_theta(xt) but not on x_{t-delta}, although the Markovian approximation in Eq. (2) uses p(x_{t+delta} | x_t, x_{t-delta}). Please clarify whether x_{t-delta} is an additional network input or is deliberately omitted.
  2. [Section 3.4] In the sentence 'Hence, CRPS it is a representative metric', the word 'it' should be removed.
  3. [Figure 16 caption] The caption contains the typo 'Spreak-skill' for 'Spread-skill'.
  4. [Section 3.3] The phrase 'spread-skill ratio 3.4' should read 'spread-skill ratio (Section 3.4)'.
  5. [Abstract; Section 6] The Abstract and Section 1 state that the code 'will be open source' and 'will be released'; since the GitHub repository is cited as evidence of reproducibility, please clarify in the final version whether the code, model weights, and evaluation pipeline are actually available at the time of publication.

Circularity Check

1 steps flagged · score 4.0 of 10

Spread-skill validation is a tuned diagnostic using test-year overfitting; core CRPS/Brier claims are independent.

  1. fitted input called prediction [Section 3.3 ('Overcoming difficulties arising from residual modeling'), Figure 3 caption, and Section 4.3 (Figures 9 and 12)]
    "Additionally, test RMSE (year 2020) are shown in dotted lines, which shows some overfitting compared to the scores in 2018 (last year in train set). ... In the experiments, we use a noise scaling coefficient of 1.05, which roughly corresponds to the percentage of overfitting observed for our underlying deterministic models. ... noise scaling has no effect on RMSE, CRPS or BrierScore, but helps to improve dispersiveness. ... ArchesWeatherGen is very slightly underdispersive, similar to IFS ENS."

    The abstract's 'true stochastic emulator' claim is supported in the evaluation by the spread-skill ratio, but the noise-scaling coefficient of 1.05 is chosen from the overfitting gap observed on the 2020 test set (Figure 3), and the ablation shows this knob acts specifically on dispersion while leaving proper scores unchanged. Consequently, the near-1 spread-skill ratio reported on 2020 is a calibrated outcome rather than an independent prediction of the generative model: the dispersion metric is matched by construction using test-year information. The fCRPS, fBrier, fES, and Ensemble-Mean-RMSE comparisons do not depend on this knob and remain independent evidence, so the circularity is partial.

full rationale

The residual decomposition xt+delta = E[xt+delta|xt] + rt+delta follows from the MSE-optimality of the deterministic model (Eqs. 4-5), and the flow-matching model is trained on held-out ERA5 residuals, so the central probabilistic skill comparisons are not circular. No load-bearing self-citation was found: the citation of Lessig et al. (2023) is supporting context, not the basis of the architecture or the main results. The main circularity-adjacent issue is the spread-skill validation: the 1.05 noise scale is selected from the 2020 test overfitting signal and OOD fine-tuning on 2019 is used to compensate for deterministic-model overfitting, making the resulting 'similar spread-skill ratio as IFS ENS' a tuned diagnostic rather than a free prediction. Separately, the abstract's claim of surpassing IFS ENS and NeuralGCM 'on all WeatherBench headline variables' is contradicted by the paper's own Figure 10 (T2m worse than IFS ENS after 9 days; NeuralGCM better on geopotential between 3 and 10 days), but this is an internal overclaim, not a circular derivation.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim relies on a standard decomposition of weather states into a conditional mean plus residual, but the dispersion correction uses a manually chosen noise scaling coefficient and an OOD fine-tuning year. These are the main free choices. No new physical entities are postulated.

free parameters (2)
  • Noise scaling coefficient rho = 1.05
    Chosen to roughly match the observed overfitting percentage of the deterministic models on the test set; used to increase input noise variance at sampling to fix underdispersion (Section 3.3).
  • OOD fine-tuning year = 2019
    Selected as out-of-distribution year adjacent to the test year 2020; fine-tuning on it improves test metrics (Section 4.4). This is a hand-chosen data split rather than a numeric fit.
assumptions (5)
  • domain assumption The conditional transition distribution p(xt+delta|xt) is narrowly centered around its expectation, so residuals have small variance.
    Used to justify the decomposition in Eq. (4) and the two-stage training; relies on near-determinism of the atmosphere at 24h lead time.
  • domain assumption A deterministic model ftheta trained with MSE loss optimally approximates E[xt+delta|xt].
    This is a standard property of MSE, but the neural network reaching the optimum is an assumption; used in Eq. (5).
  • domain assumption The Markovian factorization in Eq. (2) holds for autoregressive sampling.
    p((xt+k delta) for k=1..K | xt, xt-delta) is approximated as a product of one-step transition distributions; without this, autoregressive generation is invalid.
  • ad hoc to paper Noise scaling by 1.05 yields the correct marginal variance of generated residuals.
    The 1.05 scaling is an ad hoc heuristic to correct underdispersion, not derived from first principles.
  • ad hoc to paper Data from 2019 is representative of test-time residuals for OOD fine-tuning.
    Assumes that the residual distribution on a year adjacent to the test set matches the test distribution after fine-tuning; this is a modeling choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ArchesWeather & ArchesWeatherGen: a deterministic and generative model for efficient ML weather forecasting." pith.science (2026). https://pith.science/paper/W2VGJJXL

@misc{pith2026241212971,
  author       = {Pith},
  title        = {Pith review of: ArchesWeather & ArchesWeatherGen: a deterministic and generative model for efficient ML weather forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W2VGJJXL}},
  note         = {Machine review of arXiv:2412.12971}
}
abstract

Weather forecasting plays a vital role in today's society, from agriculture and logistics to predicting the output of renewable energies, and preparing for extreme weather events. Deep learning weather forecasting models trained with the next state prediction objective on ERA5 have shown great success compared to numerical global circulation models. However, for a wide range of applications, being able to provide representative samples from the distribution of possible future weather states is critical. In this paper, we propose a methodology to leverage deterministic weather models in the design of probabilistic weather models, leading to improved performance and reduced computing costs. We first introduce \textbf{ArchesWeather}, a transformer-based deterministic model that improves upon Pangu-Weather by removing overrestrictive inductive priors. We then design a probabilistic weather model called \textbf{ArchesWeatherGen} based on flow matching, a modern variant of diffusion models, that is trained to project ArchesWeather's predictions to the distribution of ERA5 weather states. ArchesWeatherGen is a true stochastic emulator of ERA5 and surpasses IFS ENS and NeuralGCM on all WeatherBench headline variables (except for NeuralGCM's geopotential). Our work also aims to democratize the use of deterministic and generative machine learning models in weather forecasting research, with academic computing resources. All models are trained at 1.5{\deg} resolution, with a training budget of $\sim$9 V100 days for ArchesWeather and $\sim$45 V100 days for ArchesWeatherGen. For inference, ArchesWeatherGen generates 15-day weather trajectories at a rate of 1 minute per ensemble member on a A100 GPU card. To make our work fully reproducible, our code and models are open source, including the complete pipeline for data preparation, training, and evaluation, at https://github.com/INRIA/geoarches .

Figures

Figures reproduced from arXiv: 2412.12971 by the authors.

Figure 1
Figure 1. Summary evaluation metrics (higher is better) on key upper air variables (Z500, Q700, T850, U850 and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visual comparison of attention schemes used in FuXi/Stormer (left), Pangu-Weather (middle) versus ours [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Geopotential (left) and wind speed (right) RMSE of a model without multi-step fine-tuning, for each year in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Main overview of our training pipeline of ArchesWeatherGen. (1) We train four ArchesWeather models by [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: RMSE skill scores of weather models for lead times up to 10 days. Models that don’t use ensembling are [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Spherical harmonics amplitude ratio for different models. For each model and wavelength, the energy at that [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Activity (standard deviation of the climatology-removed forecast across spatial locations) of our models for [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Ensemble Mean RMSE skill scores of ArchesWeatherGen, compared to our small deterministic ensemble [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Evaluation of ArchesWeatherGen compared to other weather ensemble models. The scores are shown as [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: fCRPS skill scores for headline physical variables in WeatherBench, comparing our models to NeuralGCM [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Rank Histogram of ArchesWeatherGen compared to NeuralGCM and IFS ENS. The rank histograms [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Ablation of our strategies to improve ensemble metrics and dispersion of residual generative models. OOD [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: Impact of the choice of deterministic model used in residual modelling on ensemble metrics. Better [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: Visualization of 700hPa specific humidity predictions made by our generative weather model on Hurricane [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]
Figure 15
Figure 15. Figure 15: fCRPS scores per headline variable [PITH_FULL_IMAGE:figures/full_fig_p026_15.png]
Figure 16
Figure 16. Figure 16: Spreak-skill ratios of our methods with different variables. Using OOD and noise scaling helps to recover [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]
Figure 17
Figure 17. Figure 17: 25-member 7-day forecast generated by ArchesWeatherGen on Hurricane Teddy, initialized September 14th, [PITH_FULL_IMAGE:figures/full_fig_p028_17.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 10 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Timestep-Conditioned Transformers for Global Weather Forecasting

    cs.LG 2026-08 conditional novelty 7.0 of 10

    A single weather transformer with inference-time timestep conditioning matches specialist models across 1 to 24 hour steps and improves rollout stability under mixed-timestep training.

  2. AI-boosted rare event sampling to characterize extreme weather

    physics.ao-ph 2025-10 conditional novelty 7.0 of 10

    AI+RES uses AI weather-forecast ensembles as a guide for rare-event simulation, yielding accurate return-period statistics for 1-in-50,000-year heatwaves at roughly 100× lower computational cost.

  3. Skillful joint probabilistic weather forecasting from marginals

    cs.LG 2025-06 conditional novelty 7.0 of 10

    FGN, a neural weather model trained only on per-location forecast scores, produces more accurate global ensemble forecasts than GenCast and captures realistic spatial correlations.

  4. Fixing the Double Penalty in Data-Driven Weather Forecasting Through a Modified Spherical Harmonic Loss Function

    cs.LG 2025-01 conditional novelty 7.0 of 10

    A parameter-free spectral loss that separates amplitude from coherence errors, applied as a fine-tuning loss to GraphCast, produces sharper deterministic weather forecasts and better ensemble spread.

  5. Weather Emulators at the Frontier of Heat Extremes Predictability

    physics.ao-ph 2026-07 accept novelty 6.0 of 10

    At 10–15 day leads, AI weather emulators can match or beat dynamical models on global temperature skill but under-represent heat-extreme intensity and lose to IFS on recall.

  6. Scaling Laws of Global Weather Models

    cs.LG 2026-02 conditional novelty 6.0 of 10

    Across five global weather models, validation loss follows power-law scaling, with wider architectures and larger training datasets outperforming deeper or smaller-data configurations.

  7. ArchesClimate: Probabilistic Decadal Ensemble Generation With Flow Matching

    physics.ao-ph 2025-09 conditional novelty 6.0 of 10

    ArchesClimate autoregressively generates 10-year monthly climate states with a flow-matching emulator trained on IPSL-CM6A-LR decadal hindcasts, producing ensembles that match the climate model for some variables.

  8. LaDCast: A Latent Diffusion Model for Medium-Range Ensemble Weather Forecasting

    cs.LG 2025-06 conditional novelty 6.0 of 10

    A latent diffusion model generates medium-range global weather ensembles at 1.5 degrees that match ECMWF IFS-ENS deterministic skill at lower compute, with weaker probabilistic spread and anecdotal cyclone advantages.

  9. Generating time-consistent dynamics with discriminator-guided image diffusion models

    cs.LG 2025-05 conditional novelty 6.0 of 10

    A time-consistency discriminator guides a pretrained image diffusion model at inference time to generate realistic, stable spatiotemporal sequences without finetuning the diffusion model.

  10. Diffusion-LAM: Probabilistic Limited Area Weather Forecasting with Diffusion

    cs.LG 2025-02 conditional novelty 6.0 of 10

    A diffusion-based limited area weather model that uses future boundary conditions from a global model improves short-lead forecast accuracy and boundary consistency.

Reference graph

Works this paper leans on

37 extracted references · 7 canonical work pages · cited by 10 Pith papers

  1. [1]

    Addison, H., Kendon, E., Ravuri, S., Aitchison, L., and Watson, P. A. Machine learning emulation of precipitation from km-scale regional climate simulations using a diffusion model. arXiv preprint arXiv:2407.14158,

    arXiv

  2. [5]

    Swinrdm: integrate swinrnn with diffusion model towards high- resolution and high-quality weather forecasting

    Chen, L., Du, F., Hu, Y ., Wang, Z., and Wang, F. Swinrdm: integrate swinrnn with diffusion model towards high- resolution and high-quality weather forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 322–330, 2023a. Chen, L., Zhong, X., Zhang, F., Cheng, Y ., Xu, Y ., Qi, Y ., and Li, H. Fuxi: a cascade machine le...

    arXiv

  3. [7]

    Scaling spherical cnns

    Esteves, C., Slotine, J.-J., and Makadia, A. Scaling spherical cnns. arXiv preprint arXiv:2306.05420,

    arXiv

  4. [8]

    S., Durand, C., Farchi, A., Bocquet, M., and Brajard, J

    Finn, T. S., Durand, C., Farchi, A., Bocquet, M., and Brajard, J. Towards diffusion models for large-scale sea-ice modelling. arXiv preprint arXiv:2406.18417,

    arXiv

  5. [10]

    A generative diffusion model for probabilistic ensembles of precipitation maps conditioned on multisensor satellite observations

    Guilloteau, C., Kerrigan, G., Nelson, K., Migliorini, G., Smyth, P., Li, R., and Foufoula-Georgiou, E. A generative diffusion model for probabilistic ensembles of precipitation maps conditioned on multisensor satellite observations. arXiv preprint arXiv:2409.16319,

    arXiv

  6. [11]

    FourCastNeXt: Optimizing FourCastNet Training for Limited Compute

    Guo, E., Ahmed, M., Sun, Y ., Mahendru, R., Yang, R., Cook, H., Leeuwenburg, T., and Evans, B. Fourcastnext: Improving fourcastnet training with limited compute. arXiv preprint arXiv:2401.05584,

    work page Pith review arXiv

  7. [13]

    The era5 global reanalysis

    Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D., et al. The era5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730): 1999–2049,

    work page 1999

  8. [15]

    Huang, C., Mu, P., Bai, C., and Watson, P. A. Tcp-diffusion: A multi-modal diffusion model for global tropical cyclone precipitation forecasting with change awareness. arXiv preprint arXiv:2410.13175, 2024a. Huang, L., Gianinazzi, L., Yu, Y ., Dueben, P. D., and Hoefler, T. Diffda: a diffusion model for weather-scale data assimilation. arXiv preprint arXi...

    arXiv

Show all 37 references
  1. [19]

    Atmorep: A stochastic model of atmosphere dynamics using large scale representation learning

    Lessig, C., Luise, I., Gong, B., Langguth, M., Stadler, S., and Schultz, M. Atmorep: A stochastic model of atmosphere dynamics using large scale representation learning. arXiv preprint arXiv:2308.13280,

    arXiv

  2. [20]

    Seeds: Emulation of weather forecast ensembles with diffusion models

    Li, L., Carver, R., Lopez-Gomez, I., Sha, F., and Anderson, J. Seeds: Emulation of weather forecast ensembles with diffusion models. arXiv preprint arXiv:2306.14066,

    arXiv

  3. [21]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y ., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    arXiv

  4. [22]

    Y ., Zepeda-Núñez, L., Schneider, T., Anderson, J., and Sha, F

    Lopez-Gomez, I., Wan, Z. Y ., Zepeda-Núñez, L., Schneider, T., Anderson, J., and Sha, F. Dynamical-generative downscaling of climate model ensembles. arXiv preprint arXiv:2410.01776,

    arXiv

  5. [23]

    Generative data assimilation of sparse weather station observations at kilometer scales

    Manshausen, P., Cohen, Y ., Pathak, J., Pritchard, M., Garg, P., Mardani, M., Kashinath, K., Byrne, S., and Brenowitz, N. Generative data assimilation of sparse weather station observations at kilometer scales. arXiv preprint arXiv:2406.16947,

    arXiv

  6. [24]

    Residual diffusion modeling for km-scale atmospheric downscaling

    Mardani, M., Brenowitz, N., Cohen, Y ., Pathak, J., Chen, C.-Y ., Liu, C.-C., Vahdat, A., Kashinath, K., Kautz, J., and Pritchard, M. Residual diffusion modeling for km-scale atmospheric downscaling. 2024a. Mardani, M., Brenowitz, N., Cohen, Y ., Pathak, J., Chen, C.-Y ., Liu,...

    arXiv

  7. [25]

    Scaling transformer neural networks for skillful and reliable medium-range weather forecasting

    Nguyen, T., Shah, R., Bansal, H., Arcomano, T., Madireddy, S., Maulik, R., Kotamarthi, V ., Foster, I., and Grover, A. Scaling transformer neural networks for skillful and reliable medium-range weather forecasting. arXiv preprint arXiv:2312.03876,

    arXiv

  8. [26]

    Heavy-tailed diffusion models

    Pandey, K., Pathak, J., Xu, Y ., Mandt, S., Pritchard, M., Vahdat, A., and Mardani, M. Heavy-tailed diffusion models. arXiv preprint arXiv:2410.14171,

    arXiv

  9. [27]

    Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators

    23 ArchesWeather & ArchesWeatherGen A PREPRINT Pathak, J., Subramanian, S., Harrington, P., Raja, S., Chattopadhyay, A., Mardani, M., Kurth, T., Hall, D., Li, Z., Azizzadenesheli, K., et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier...

    arXiv

  10. [29]

    Weatherbench 2: A benchmark for the next generation of data-driven global weather models

    Rasp, S., Hoyer, S., Merose, A., Langmore, I., Battaglia, P., Russel, T., Sanchez-Gonzalez, A., Yang, V ., Carver, R., Agrawal, S., et al. Weatherbench 2: A benchmark for the next generation of data-driven global weather models. arXiv preprint arXiv:2308.15560,

    arXiv

  11. [30]

    Glu variants improve transformer

    Shazeer, N. Glu variants improve transformer. arXiv preprint arXiv:2002.05202,

    2002 arXiv

  12. [32]

    Probabilistic precipitation downscaling with optical flow-guided diffusion

    Srivastava, P., Yang, R., Kerrigan, G., Dresdner, G., McGibbon, J., Bretherton, C., and Mandt, S. Probabilistic precipitation downscaling with optical flow-guided diffusion. arXiv preprint arXiv:2312.06071,

    arXiv

  13. [33]

    Workshop on Predictability

    Talagrand, O. Evaluation of probabilistic prediction systems. In Workshop Proceedings" Workshop on Predictability", 20-22 October 1997, ECMWF , Reading, UK,

    1997

  14. [35]

    Watt, R. A. and Mansfield, L. A. Generative diffusion-based downscaling for climate. arXiv preprint arXiv:2404.17752,

    arXiv

  15. [36]

    Debias the training of diffusion models

    Yu, H., Shen, L., Huang, J., Zhou, M., Li, H., and Zhao, F. Debias the training of diffusion models. arXiv preprint arXiv:2310.08442,

    arXiv

  16. [37]

    URL https://www.ecmwf

    doi:10.21957/8b50609a0f. URL https://www.ecmwf. int/en/about/media-centre/aifs-blog/2024/accuracy-versus-activity . 25 ArchesWeather & ArchesWeatherGen A PREPRINT A Appendix: Additional details A.1 Training details We denote (xt)t∈D the historical trajectory of ERA5, indexed b...

    2024 doi

  17. [1994]

    Graphcast: Learning skillful medium-range global weather forecasting

    Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton- Rosen, Z., Hu, W., et al. Graphcast: Learning skillful medium-range global weather forecasting. arXiv preprint arXiv:2212.12794,

    arXiv

  18. [1999]

    Can ai be enabled to dynamical downscaling? a latent diffusion model to mimic km-scale cosmo5

    Tomasi, E., Franch, G., and Cristoforetti, M. Can ai be enabled to dynamical downscaling? a latent diffusion model to mimic km-scale cosmo5. 0_clm9 simulations. EGUsphere, 2024:1,

    2024

  19. [2000]

    arXiv preprint arXiv:2408.12632,

    arXiv

  20. [2007]

    Cascast: Skillful high-resolution precipitation nowcasting via cascaded modelling

    Gong, J., Bai, L., Ye, P., Xu, W., Liu, N., Dai, J., Yang, X., and Ouyang, W. Cascast: Skillful high-resolution precipitation nowcasting via cascaded modelling. arXiv preprint arXiv:2402.04290,

    arXiv

  21. [2014]

    Neural general circulation models

    Kochkov, D., Yuval, J., Langmore, I., Norgaard, P., Smith, J., Mooers, G., Lottes, J., Rasp, S., Düben, P., Klöwer, M., et al. Neural general circulation models. arXiv preprint arXiv:2311.07222,

    arXiv

  22. [2015]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502,

    2010 arXiv

  23. [2017]

    Deep learning for day forecasts from sparse observations

    Andrychowicz, M., Espeholt, L., Li, D., Merchant, S., Merose, A., Zyda, F., Agrawal, S., and Kalchbrenner, N. Deep learning for day forecasts from sparse observations. arXiv preprint arXiv:2306.06079,

    arXiv

  24. [2018]

    Gencast: Diffusion-based ensemble forecasting for medium-range weather

    Price, I., Sanchez-Gonzalez, A., Alet, F., Ewalds, T., El-Kadi, A., Stott, J., Mohamed, S., Battaglia, P., Lam, R., and Willson, M. Gencast: Diffusion-based ensemble forecasting for medium-range weather. arXiv preprint arXiv:2312.15796,

    arXiv

  25. [2020]

    Axial attention in multidimensional transformers

    22 ArchesWeather & ArchesWeatherGen A PREPRINT Ho, J., Kalchbrenner, N., Weissenborn, D., and Salimans, T. Axial attention in multidimensional transformers. arXiv preprint arXiv:1912.12180,

    1912 arXiv

  26. [2021]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929,

    2010 arXiv

  27. [2022]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

    arXiv

  28. [2023]

    Pangu-weather: A 3d high-resolution model for fast and accurate global weather forecast

    Bi, K., Xie, L., Zhang, H., Chen, X., Gu, X., and Tian, Q. Pangu-weather: A 3d high-resolution model for fast and accurate global weather forecast. arXiv preprint arXiv:2211.02556,

    arXiv

  29. [2024]

    Checkerboard artifact free sub-pixel convolution: A note on sub-pixel convolution, resize convolution and convolution resize

    Aitken, A., Ledig, C., Theis, L., Caballero, J., Wang, Z., and Shi, W. Checkerboard artifact free sub-pixel convolution: A note on sub-pixel convolution, resize convolution and convolution resize. arXiv preprint arXiv:1707.02937,

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.