Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

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

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A latent-space diffusion transformer approaches the operational IFS-ENS ensemble's medium-range skill while running a 50-member forecast in under a minute.

desk verdict LaDCast is a credible new latent-diffusion approach to global ensemble forecasting with solid deterministic skill, but the headline claim of superior cyclone tracking is not yet supported by the evidence. read the letter →

arxiv 2506.09193 v1 pith:QNFOHOS7 submitted 2025-06-10 cs.LG

classification cs.LG
keywords latentdiffusionmodelmedium-rangeensembleforecastingprobabilisticweatherERA5neuralcompressionGeoRoPEcyclonetrackingtransformer
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

LaDCast is a machine-learning weather model that produces 15-day global ensemble forecasts entirely in a learned latent space rather than on the physical grid. The paper argues that a transformer-based diffusion model, conditioned on four compressed snapshots of ERA5, can reach deterministic and probabilistic skill close to the operational ensemble IFS-ENS without explicitly perturbing the initial state, and that it tracks rare storms more accurately: in case studies of Typhoon Kong-rey, Hurricane Dorian, and Hurricane Lorenzo, the LaDCast ensemble follows the observed paths better than IFS-ENS does. This matters because operational ensembles are expensive and locked to 12-hour cycles, whereas LaDCast generates a 50-member forecast in under a minute on a single GPU and can initialize at arbitrary hours, opening a route to kilometer-scale real-time probabilistic forecasting. The paper's stated limitation is that the autoencoder's reconstruction error caps forecast accuracy, so every downstream result inherits whatever the latent compression discards.

What carries the argument

The load-bearing mechanism is the compression–diffusion pair: a Deep Compression Autoencoder (DC-AE) maps ERA5 snapshots between physical fields and an 84×15×30 latent tensor, and an EDM-style diffusion transformer denoises the next latent block given the previous four, using GeoRoPE (rotary position embeddings adapted to longitude periodicity and latitude circulation bands) and a dual-stream attention block that encodes conditioning and target tokens separately. A sinusoidal year-progress embedding supplies seasonality, and a deterministic probability-flow ODE sampled with twenty Heun steps generates the ensemble through the Gaussian noise initialization. The DC-AE is the part that sets the ceiling: information lost at 64× spatial compression cannot be recovered by the diffusion stage, which is why the paper names reconstruction error as the limiting factor.

What would settle it

Encode and decode the verifying ERA5 fields for 2018 with the DC-AE, then compare that reconstruction RMSE with LaDCast's forecast RMSE at each lead time; if the reconstructed truth is no more accurate than IFS-ENS at any lead where LaDCast claims parity, the autoencoder rather than the diffusion model is carrying the result.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that useful ensemble uncertainty can be generated by sampling in a learned latent space instead of perturbing initial conditions. Trained on ERA5 (1979–2017) at 1.5° resolution, the 1.6B-parameter LaDCast produces 50-member, 15-day forecasts whose ensemble-mean RMSE approaches IFS-ENS at extended lead times, whose 6-hour precipitation RMSE is lower than IFS-ENS, and whose CRPS is close to IFS-ENS despite showing smaller spread because the conditioning initial profile is unperturbed. The cyclone case studies are presented as the strongest result: in all three, the LaDCast ensemble reproduces the IBTrACS-observed path more consistently than the IFS-ENS ensemble, even though the Dorian ensemble-mean track is misled by a heuristic tracker latching onto a spurious low over Florida. The paper also reports that doing the computation on an 84×15×30 latent code cuts storage and compute by orders of magnitude.

Load-bearing premise

The entire argument rests on the assumption that compressing the atmosphere into an 84-channel latent tensor leaves the week-two forecast skill and the storm-track structure intact; the paper's own limitation statement shows that this is exactly where accuracy is currently lost.

Editorial extensions

If this is right

  • If the central claim holds, LaDCast's 50-member, 15-day global forecast at 1.5° takes under a minute on one GPU, putting operational-grade ensemble forecasting within reach of institutions that cannot run a full NWP center.
  • If the central claim holds, arbitrary-hour initialization removes the fixed 12-hour cycle of the operational ensemble, so forecasts can be re-issued whenever a storm reorganizes.
  • If the central claim holds, the three out-of-training cyclone cases (Kong-rey, Dorian, Lorenzo) indicate that unperturbed conditional generation can give storm-track ensembles with less misleading spread than the operational perturbed ensemble, supporting the paper's disaster-preparedness claim.
  • If the central claim holds, moving the heavy computation into the compressed latent space is the paper's stated path toward kilometer-scale, real-time global probabilistic forecasting.

Reading between the lines

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

  • Beyond the paper: the reconstruction-RMSE ceiling is separable from the diffusion forecast, so subtracting the DC-AE's reconstruction error from LaDCast's forecast RMSE at each lead time would reveal the diffusion model's true contribution and make the compression bottleneck quantitative.
  • Beyond the paper: the paper finds that perturbing latent initial conditions does not help overall; a next test is to perturb only the least-reconstructed latent channels, since the DC-AE's errors concentrate in particular variables, to see whether targeted perturbation can close the CRPS gap without weakening the cyclone tracks.
  • Beyond the paper: because all sampling happens in one fixed latent code, the same DC-AE could be coupled with cheaper generative samplers; if the storm-track advantage survives that swap, it is a property of the latent representation rather than of the diffusion recipe.
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

3 major / 6 minor

Summary. The manuscript introduces LaDCast, a latent diffusion model for global medium-range ensemble weather forecasting. An autoencoder compresses ERA5 fields at 1.5° into an 84×15×30 latent space, and a transformer-based diffusion model with GeoRoPE and dual-stream attention generates sequential forecasts in that space. The authors train on 1979–2017 and evaluate 2018, with a 2019 appendix check, using WeatherBench2 latitude-weighted RMSE and CRPS. They compare 50-member LaDCast ensembles with IFS-ENS and deterministic ML baselines, reporting deterministic skill comparable to IFS-ENS at longer lead times, a CRPS deficit attributed to underdispersion, and three cyclone case studies (Kong-rey, Dorian, Lorenzo) interpreted as superior trajectory tracking. They also claim large reductions in storage and compute and open-source the code and models.

Significance. If the results hold, LaDCast would be a notable advance: it is the first global latent-diffusion weather forecasting framework, demonstrates that a deeply compressed latent space can support 15-day ensemble rollouts, and shows a credible path toward cheaper probabilistic forecasting. The paper deserves credit for using standard WeatherBench2 metrics, testing on a held-out year with a second-year check, running controlled ablations (model size, ensemble size, sequence length, sampling steps), and releasing code and models. The evaluation is not circular: training targets ERA5, and comparisons are made against IFS-ENS and IBTrACS. However, the probabilistic-skill claim is weakened by the reported CRPS deficit, and the extreme-event claim rests on three hand-selected storms without quantitative track-error statistics. The significance of the paper would be substantially higher if these two pillars were supported with quantitative evidence.

major comments (3)
  1. [Section 5, Figure 6; Abstract] The abstract's claim of 'probabilistic skill close to that of IFS-ENS' is not supported by the reported CRPS. The text states that LaDCast 'has a smaller spread than IFS-ENS, which results in a higher CRPS value', but no numerical CRPS values, lead-time breakdowns, or uncertainty estimates are provided. Since spread and reliability are the core of probabilistic skill, the paper should report CRPS by variable and lead time together with calibration diagnostics (e.g., rank histograms, spread-error ratio, or a CRPS decomposition). The authors should either quantify how close the CRPS gap actually is or soften the probabilistic wording in the abstract.
  2. [Section 5 and Appendix G] The claim of 'superior performance in tracking rare extreme events such as cyclones' is not yet established. The evidence is limited to three retrospective storms, with storm-specific tracker settings: Appendix G gives search boxes [7,4,1] for Lorenzo, [7,5,1] for Kong-rey, and [6,3,0] for Dorian, and enables the Z700 fallback and land-sea mask only for Dorian. Figures 2, 7, and 8 are visual trajectory overlays with no along-track/cross-track error, no mean absolute track error versus lead time, and no spread-skill analysis. The authors should provide quantitative track-error metrics, preferably over a larger pre-specified storm set, or withdraw the word 'superior' from the abstract and conclusions.
  3. [Section 6, Table 7, and Figure 10] The acknowledged autoencoder bottleneck is not quantified relative to the forecast skill claims. Section 6 states that the model is 'constrained by the reconstruction error of its deep-compression autoencoder, which limits forecasting accuracy', and Table 7 shows substantial per-variable reconstruction RMSE (e.g., mean sea level pressure around 30 hPa). Since all forecasts are generated and decoded in this latent space, the reported deterministic and probabilistic errors include this compression error. To separate diffusion error from reconstruction error, the authors should report a 'reconstruction ceiling'—for instance, the RMSE and CRPS obtained by decoding the true latent trajectory—and compare it with LaDCast's forecast error. Without this, the headline 'close to IFS-ENS' cannot be cleanly attributed to the forecasting model itself.
minor comments (6)
  1. [Section 3, Eq. (2)] The denoising target is written as x0 inside the norm, while the surrounding notation uses zout; please align the notation throughout the equation.
  2. [Table 1] Please state the evaluation periods explicitly in the 'Year' column and clarify whether the spatial compression ratios are defined over the spatial dimensions only, since the VAEformer and DC-AE comparisons use different latent geometries.
  3. [Section 4] The text says the model selects 'six single variables, six atmospheric variables and five static features', which sums to 17 channels, while Table 6 lists these plus a clock feature; clarify how the clock feature enters the 89 input channels and how static features are combined.
  4. [Appendix D.1] The text says the 1.5° grid is '121×240' and then 'adjusted to an even latitude dimension by cropping the south pole'; state the final grid size explicitly (120×240) to avoid ambiguity.
  5. [Figure 5] The caption mentions a 3-day forecast during Hurricane Dorian, but the surrounding text refers to total precipitation RMSE for 2018; clarify whether the visualization is a model-output example or an RMSE map and specify the lead time.
  6. [Section 5] The text states that the training data cover 1979–2017 and that all three storms are outside the training range; this is correct, but please state explicitly that Hurricane Dorian and Lorenzo are 2019 storms so readers do not conflate the 2018 evaluation year with the case-study years.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LaDCast's central claims are supported by external ERA5, IFS-ENS, and IBTrACS benchmarks, and no prediction is defined by or fitted to its evaluation target.

full rationale

The derivation chain is self-contained. The DC-AE is trained to minimize a relative L2 reconstruction error on ERA5 snapshots (Section 3), and the diffusion model is trained on the resulting latent sequences with an EDM denoising objective (Eqs. 2-4). The evaluation compares against ERA5 reanalysis, IFS-ENS, and IBTrACS, all of which are external to the training objective. The ensemble spread arises from sampling diffusion noise in latent space, not from perturbations fitted to the benchmark, so the claim 'without any explicit perturbations' is not a fitted input renamed as a prediction. The cyclone case studies are the weakest part of the paper: they involve three hand-selected storms, per-storm tracker parameters in Appendix G, visual trajectory overlays, and no quantitative track-error metric. However, this is an evidence-strength and correctness limitation, not circularity, because the tracker is applied symmetrically to model outputs and no tracked quantity is defined in terms of LaDCast's outputs or fitted to the benchmark. The paper's own limitation statement (Section 6) concedes that 'The current LaDCast model is constrained by the reconstruction error of its deep-compression autoencoder, which limits forecasting accuracy'; this weakens the headline claims but does not make the derivation circular. Self-citations [20] and [51] are background references for variance-exploding diffusion formulations and physical-consistency losses; they are not load-bearing for the main forecast-skill or cyclone-tracking claims. No equation equates a predicted quantity to its input, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

This is an empirical ML system, so the ledger is dominated by design choices and data assumptions rather than by physical postulates. The most load-bearing items are latent-code sufficiency, fairness of the IFS-ENS comparison, and adequacy of diffusion sampling for uncertainty. The per-storm tracker tuning is the clearest instance of a fitted, ad hoc choice affecting a headline claim.

free parameters (4)
  • Cyclone tracker search boxes and field switches = [7,4,1] for Lorenzo; [7,5,1] for Kong-rey; [6,3,0] for Dorian; Z700 and land-sea mask used only for Dorian
    These per-storm choices tune the post-processing that produces the headline claim of superior cyclone tracking. They are not fixed in advance and are not applied uniformly across cases.
  • Latent bottleneck geometry = 84 latent channels at 15x30 spatial resolution
    The DC-AE output dimensionality is a design choice that determines compression ratio and what information the diffusion model can access. No experiment shows it is optimal or sufficient.
  • Diffusion sampling schedule = sigma_max=80, sigma_min=0.002, rho=7, 20 Heun steps
    Hand-selected EDM defaults and step count. The ablation varies only the step count, not the noise schedule or solver details.
  • Model capacity and output sequence length = 1.6B or 375M parameters; default 1-to-4 sequence
    Architecture size and 1-to-4 rollout length are chosen after ablations show they matter. They are not derived from first principles and may not transfer to other resolutions.
assumptions (6)
  • domain assumption ERA5 reanalysis approximates atmospheric truth.
    All training targets and verification metrics treat ERA5 as ground truth. Any systematic reanalysis error is inherited by the model and by the skill scores. Invoked throughout Section 5.
  • domain assumption Evaluation against WB2-hosted IFS-ENS is a fair comparison.
    Skill claims are relative to the 1.5-degree IFS-ENS baseline hosted by WeatherBench2. If the operational IFS-ENS runs at much higher resolution, the comparison is not fully apples-to-apples. Invoked in Section 5.
  • domain assumption The latent code preserves enough information for forecasting.
    The entire framework predicts in latent space, so reconstruction fidelity bounds forecast quality. The paper's own limitations section says this is the main constraint. Invoked in Section 6.
  • domain assumption Diffusion sampling stochasticity adequately represents forecast uncertainty without perturbing initial conditions.
    The paper assumes ensemble spread from sampling is sufficient. Its own CRPS results show spread is lower than IFS-ENS, so this assumption is only partially satisfied. Invoked in Section 5 and Figure 6.
  • domain assumption The heuristic cyclone tracker identifies true storm centers.
    Extreme-event claims rest on local-minima tracking in MSLP and Z700 with per-storm settings. Tracker errors can create artificial skill differences. Invoked in Appendix G.
  • standard math Standard EDM and score-based diffusion theory is correct.
    The preconditioning, probability-flow ODE, and Heun solver are taken from Karras et al. and Song et al. without re-derivation. Invoked in Section 3 and Appendix B.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LaDCast: A Latent Diffusion Model for Medium-Range Ensemble Weather Forecasting." pith.science (2026). https://pith.science/paper/QNFOHOS7

@misc{pith2026250609193,
  author       = {Pith},
  title        = {Pith review of: LaDCast: A Latent Diffusion Model for Medium-Range Ensemble Weather Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QNFOHOS7}},
  note         = {Machine review of arXiv:2506.09193}
}
read the original abstract

Accurate probabilistic weather forecasting demands both high accuracy and efficient uncertainty quantification, challenges that overburden both ensemble numerical weather prediction (NWP) and recent machine-learning methods. We introduce LaDCast, the first global latent-diffusion framework for medium-range ensemble forecasting, which generates hourly ensemble forecasts entirely in a learned latent space. An autoencoder compresses high-dimensional ERA5 reanalysis fields into a compact representation, and a transformer-based diffusion model produces sequential latent updates with arbitrary hour initialization. The model incorporates Geometric Rotary Position Embedding (GeoRoPE) to account for the Earth's spherical geometry, a dual-stream attention mechanism for efficient conditioning, and sinusoidal temporal embeddings to capture seasonal patterns. LaDCast achieves deterministic and probabilistic skill close to that of the European Centre for Medium-Range Forecast IFS-ENS, without any explicit perturbations. Notably, LaDCast demonstrates superior performance in tracking rare extreme events such as cyclones, capturing their trajectories more accurately than established models. By operating in latent space, LaDCast reduces storage and compute by orders of magnitude, demonstrating a practical path toward forecasting at kilometer-scale resolution in real time. We open-source our code and models and provide the training and evaluation pipelines at: https://github.com/tonyzyl/ladcast.

Figures

Figures reproduced from arXiv: 2506.09193 by the authors.

Figure 1
Figure 1. Visualization of a 5-day forecast showing the 10m wind speed forecast for the typhoon [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Fifty 10-day forecast trajectories from IFS-ENS and LaDCast for Typhoon Kong-rey [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the LaDCast model: Predictions are made in the latent space and the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Latitude-weighted RMSE for 2018. The LaDCast shows better performance than determin [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Total precipitation 6hr RMSE for 2018 compared to IFS-ENS. The visualization shows a [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Results of the CRPS(↓) for 2018, Due to the deterministic conditioning, the spread of LaDCast is lower than IFS-ENS, which results in a higher CRPS value. Lorenzo (2019). The training data covers the period 1979-2017, so all three storms lie outside the training range.…
Figure 7
Figure 7. Figure 7: 50 trajectories, 10-day forecast of IFS-ENS and LaDCast for hurricane Dorian (2019), [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: 50 trajectories, 7-day forecasts from IFS-ENS and LaDCast for Hurricane Lorenzo (2019), [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: An illustration of the spherical convolutional kernel. Kernels’ weights at the corresponding [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Latitude-weighted RMSE of LaDCast for 2018 and 2019. Error bars indicate the standard deviation computed from 240 samples. We initialize the Gaussian noise prior for all results with a seed tied to each ensemble member’s index to ensure reproducibility. In the ablatio…
Figure 11
Figure 11. Figure 11: Ablation study of the effect of input-sequence length. [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: Ablation study on scaling and ensemble size. [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: Ablation study of latent-space perturbations. [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Ablation study on the number of reverse-sampling steps. [PITH_FULL_IMAGE:figures/full_fig_p023_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Data-Efficient Ensemble Weather Forecasting with Diffusion Models

    cs.LG 2025-09 conditional novelty 4.0 of 10

    Training an autoregressive diffusion weather forecaster on 20% of ERA5 data selected uniformly by calendar month matches full-data CRPS/RMSE and improves the spread-skill ratio on the 2018 test year.

Reference graph

Works this paper leans on

54 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    Appa: Bending weather dynamics with latent diffusion models for global data assimilation.arXiv preprint arXiv:2504.18720, 2025

    Gérôme Andry, François Rozet, Sacha Lewin, Omer Rochman, Victor Mangeleer, Matthias Pirlet, Elise Faulx, Marilaure Grégoire, and Gilles Louppe. Appa: Bending weather dynamics with latent diffusion models for global data assimilation.arXiv preprint arXiv:2504.18720, 2025

  2. [2]

    Zied Ben Bouallègue, Mariana CA Clare, Linus Magnusson, Estibaliz Gascon, Michael Maier-Gerber, Martin Janoušek, Mark Rodwell, Florian Pinault, Jesper S Dramsch, Simon TK Lang, et al. The rise of data-driven weather forecasting: A first statistical assessment of machine learning–based weather forecasts in an operational-like context.Bulletin of the Americ...

  3. [3]

    Pangu-weather: A 3d high-resolution model for fast and accurate global weather forecast.arXiv preprint arXiv:2211.02556, 2022

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Pangu-weather: A 3d high-resolution model for fast and accurate global weather forecast.arXiv preprint arXiv:2211.02556, 2022

  4. [4]

    Aurora: A foundation model of the atmosphere.arXiv preprint arXiv:2405.13063, 2024

    Cristian Bodnar, Wessel P Bruinsma, Ana Lucic, Megan Stanley, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan Weyn, Haiyu Dong, Anna Vaughan, et al. Aurora: A foundation model of the atmosphere.arXiv preprint arXiv:2405.13063, 2024

  5. [5]

    Deep compression autoencoder for efficient high-resolution diffusion models.arXiv preprint arXiv:2410.10733, 2024

    Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models.arXiv preprint arXiv:2410.10733, 2024

  6. [6]

    Arch- esweather & archesweathergen: a deterministic and generative model for efficient ml weather forecasting

    Guillaume Couairon, Renu Singh, Anastase Charantonis, Christian Lessig, and Claire Monteleoni. Arch- esweather & archesweathergen: a deterministic and generative model for efficient ml weather forecasting. arXiv preprint arXiv:2412.12971, 2024

  7. [7]

    Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nature Communications, 15(1):10416, 2024

    Pan Du, Meet Hemant Parikh, Xiantao Fan, Xin-Yang Liu, and Jian-Xun Wang. Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nature Communications, 15(1):10416, 2024

  8. [8]

    Number 5

    ECMWF.IFS Documentation CY46R1 - Part V: Ensemble Prediction System. Number 5. 2019

Show all 54 references
  1. [9]

    Han Gao, Sebastian Kaltenbach, and Petros Koumoutsakos. Generative learning of the solution of parametric partial differential equations using guided diffusion models and virtual observations.Computer Methods in Applied Mechanics and Engineering, 435:117654, 2025

  2. [10]

    Prediff: Precipitation nowcasting with latent diffusion models.Advances in Neural Information Processing Systems, 36:78621–78656, 2023

    Zhihan Gao, Xingjian Shi, Boran Han, Hao Wang, Xiaoyong Jin, Danielle Maddix, Yi Zhu, Mu Li, and Yuyang Bernie Wang. Prediff: Precipitation nowcasting with latent diffusion models.Advances in Neural Information Processing Systems, 36:78621–78656, 2023

  3. [11]

    Strictly proper scoring rules, prediction, and estimation.Journal of the American statistical Association, 102(477):359–378, 2007

    Tilmann Gneiting and Adrian E Raftery. Strictly proper scoring rules, prediction, and estimation.Journal of the American statistical Association, 102(477):359–378, 2007

  4. [12]

    Accelerate: Training and inference at scale made simple, efficient and adaptable.https://github.com/huggingface/accelerate, 2022

    Sylvain Gugger, Lysandre Debut, Thomas Wolf, Philipp Schmid, Zachary Mueller, Sourab Mangrulkar, Marc Sun, and Benjamin Bossan. Accelerate: Training and inference at scale made simple, efficient and adaptable.https://github.com/huggingface/accelerate, 2022

  5. [13]

    Cra5: Extreme compression of era5 for portable global climate and weather research via an efficient variational transformer.arXiv preprint arXiv:2405.03376, 2024

    Tao Han, Zhenghao Chen, Song Guo, Wanghan Xu, and Lei Bai. Cra5: Extreme compression of era5 for portable global climate and weather research via an efficient variational transformer.arXiv preprint arXiv:2405.03376, 2024

  6. [14]

    Fengwu-ghr: Learning the kilometer-scale medium-range global weather forecasting

    Tao Han, Song Guo, Fenghua Ling, Kang Chen, Junchao Gong, Jingjia Luo, Junxia Gu, Kan Dai, Wanli Ouyang, and Lei Bai. Fengwu-ghr: Learning the kilometer-scale medium-range global weather forecasting. arXiv preprint arXiv:2402.00059, 2024

  7. [15]

    Rotary position embedding for vision transformer

    Byeongho Heo, Song Park, Dongyoon Han, and Sangdoo Yun. Rotary position embedding for vision transformer. InEuropean Conference on Computer Vision, pages 289–305. Springer, 2024

  8. [16]

    The era5 global reanalysis.Quarterly Journal of the Royal Meteorological Society, 146(730):1999–2049, 2020

    Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquín Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis.Quarterly Journal of the Royal Meteorological Society, 146(730):1999–2049, 2020

  9. [17]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

  10. [18]

    Diffusionpde: Generative pde-solving under partial observation, 2024

    Jiahe Huang, Guandao Yang, Zichen Wang, and Jeong Joon Park. Diffusionpde: Generative pde-solving under partial observation, 2024. 11

  11. [19]

    Diffda: a diffusion model for weather-scale data assimilation.arXiv preprint arXiv:2401.05932, 2024

    Langwen Huang, Lukas Gianinazzi, Yuejiang Yu, Peter D Dueben, and Torsten Hoefler. Diffda: a diffusion model for weather-scale data assimilation.arXiv preprint arXiv:2401.05932, 2024

  12. [20]

    Cocogen: Physically-consistent and conditioned score-based generative models for forward and inverse problems, 2023

    Christian Jacobsen, Yilin Zhuang, and Karthik Duraisamy. Cocogen: Physically-consistent and conditioned score-based generative models for forward and inverse problems, 2023

  13. [21]

    Insights into atmospheric predictability through global convection-permitting model simulations

    Falko Judt. Insights into atmospheric predictability through global convection-permitting model simulations. Journal of the Atmospheric Sciences, 75(5):1477–1497, 2018

  14. [22]

    Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35:26565–26577, 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35:26565–26577, 2022

  15. [23]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Diederik P Kingma. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

  16. [24]

    The international best track archive for climate stewardship (ibtracs) unifying tropical cyclone data.Bulletin of the American Meteorological Society, 91(3):363–376, 2010

    Kenneth R Knapp, Michael C Kruk, David H Levinson, Howard J Diamond, and Charles J Neumann. The international best track archive for climate stewardship (ibtracs) unifying tropical cyclone data.Bulletin of the American Meteorological Society, 91(3):363–376, 2010

  17. [25]

    Neural general circulation models for weather and climate

    Dmitrii Kochkov, Janni Yuval, Ian Langmore, Peter Norgaard, Jamie Smith, Griffin Mooers, Milan Klöwer, James Lottes, Stephan Rasp, Peter Düben, et al. Neural general circulation models for weather and climate. Nature, 632(8027):1060–1066, 2024

  18. [26]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

  19. [27]

    Flux.https://github.com/black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/black-forest-labs/flux, 2024

  20. [28]

    Graphcast: Learning skillful medium-range global weather forecasting.arXiv preprint arXiv:2212.12794, 2022

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Graphcast: Learning skillful medium-range global weather forecasting.arXiv preprint arXiv:2212.12794, 2022

  21. [29]

    Latent diffusion models for generative precipitation nowcasting with accurate uncertainty quantification.arXiv preprint arXiv:2304.12891, 2023

    Jussi Leinonen, Ulrich Hamann, Daniele Nerini, Urs Germann, and Gabriele Franch. Latent diffusion models for generative precipitation nowcasting with accurate uncertainty quantification.arXiv preprint arXiv:2304.12891, 2023

  22. [30]

    Datasets: A community library for natural language processing

    Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Ca...

  23. [31]

    The predictability of a flow which possesses many scales of motion.Tellus, 21(3):289– 307, 1969

    Edward N Lorenz. The predictability of a flow which possesses many scales of motion.Tellus, 21(3):289– 307, 1969

  24. [32]

    A new grid for the ifs.ECMWF newsletter, 146(23-28):321, 2016

    Sylvie Malardel, Nils Wedi, Willem Deconinck, Michail Diamantakis, Christian Kühnlein, George Mozdzynski, Mats Hamrud, and Piotr Smolarkiewicz. A new grid for the ifs.ECMWF newsletter, 146(23-28):321, 2016

  25. [33]

    Scaling transformer neural networks for skillful and reliable medium-range weather forecasting.Advances in Neural Information Processing Systems, 37:68740–68771, 2024

    Tung Nguyen, Rohan Shah, Hritik Bansal, Troy Arcomano, Romit Maulik, Rao Kotamarthi, Ian Foster, Sandeep Madireddy, and Aditya Grover. Scaling transformer neural networks for skillful and reliable medium-range weather forecasting.Advances in Neural Information Processing Syste...

  26. [34]

    Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators.arXiv preprint arXiv:2202.11214, 2022

    Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural oper...

  27. [35]

    Gencast: Diffusion-based ensemble forecasting for medium-range weather.arXiv preprint arXiv:2312.15796, 2023

    Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, et al. Gencast: Diffusion-based ensemble forecasting for medium-range weather.arXiv preprint arXiv:2312.15796, 2023. 12

  28. [36]

    Weatherbench 2: A benchmark for the next generation of data-driven global weather models.Journal of Advances in Modeling Earth Systems, 16(6):e2023MS004019, 2024

    Stephan Rasp, Stephan Hoyer, Alexander Merose, Ian Langmore, Peter Battaglia, Tyler Russell, Alvaro Sanchez-Gonzalez, Vivian Yang, Rob Carver, Shreya Agrawal, et al. Weatherbench 2: A benchmark for the next generation of data-driven global weather models.Journal of Advances in...

  29. [37]

    Jumpiness in ensemble forecasts of atlantic tropical cyclone tracks.Weather and Forecasting, 39(1):203–215, 2024

    David S Richardson, Hannah L Cloke, John A Methven, and Florian Pappenberger. Jumpiness in ensemble forecasts of atlantic tropical cyclone tracks.Weather and Forecasting, 39(1):203–215, 2024

  30. [38]

    High-resolution image synthesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022

  31. [39]

    Probablistic emulation of a global climate model with spherical dyffusion.Advances in Neural Information Processing Systems, 37:127610–127644, 2024

    Salva Rühling Cachay, Brian Henn, Oliver Watt-Meyer, Christopher S Bretherton, and Rose Yu. Probablistic emulation of a global climate model with spherical dyffusion.Advances in Neural Information Processing Systems, 37:127610–127644, 2024

  32. [40]

    Dyffusion: A dynamics-informed diffusion model for spatiotemporal forecasting.Advances in neural information processing systems, 36:45259–45287, 2023

    Salva Rühling Cachay, Bo Zhao, Hailey Joren, and Rose Yu. Dyffusion: A dynamics-informed diffusion model for spatiotemporal forecasting.Advances in neural information processing systems, 36:45259–45287, 2023

  33. [41]

    The general circulation of the atmosphere.Annu

    Tapio Schneider. The general circulation of the atmosphere.Annu. Rev. Earth Planet. Sci., 34(1):655–688, 2006

  34. [42]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations, 2021

  35. [43]

    Cambridge University Press, 2007

    David J Stensrud.Parameterization schemes: keys to understanding numerical weather prediction models. Cambridge University Press, 2007

  36. [44]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

  37. [45]

    Diffusers: State-of-the-art diffusion models.https://github.com/huggingface/diffusers, 2022

    Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models.https://github.com/huggingface/diffusers, 2022

  38. [46]

    306: FM92–GRIB2–11–2003: GRIB Edition2 (binary only)

    World Meteorological Organization.Manual on Codes No. 306: FM92–GRIB2–11–2003: GRIB Edition2 (binary only). Geneva, Switzerland, November 2003

  39. [47]

    Sana: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2024

    Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2024

  40. [48]

    Diffcast: A unified framework via residual diffusion for precipitation nowcasting

    Demin Yu, Xutao Li, Yunming Ye, Baoquan Zhang, Chuyao Luo, Kuai Dai, Rui Wang, and Xunlai Chen. Diffcast: A unified framework via residual diffusion for precipitation nowcasting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27758–2...

  41. [49]

    Estimation of the continuous ranked probability score with limited information and applications to ensemble weather forecasts.Mathematical Geosciences, 50(2):209–234, 2018

    Michaël Zamo and Philippe Naveau. Estimation of the continuous ranked probability score with limited information and applications to ensemble weather forecasts.Mathematical Geosciences, 50(2):209–234, 2018

  42. [50]

    Are transformers effective for time series forecasting? InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121–11128, 2023

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121–11128, 2023

  43. [51]

    Yilin Zhuang, Sibo Cheng, and Karthik Duraisamy. Spatially-aware diffusion models with cross-attention for global field reconstruction with sparse observations.Computer Methods in Applied Mechanics and Engineering, 435:117623, 2025. A Broader impact. The adoption of a latent d...

  44. [52]

    Initialization: The tracker starts with a known initial position (latitude-longitude) of a tropical cyclone at timet 0

  45. [53]

    21 Figure 13: Ablation study of latent-space perturbations

    Grid Alignment: All positional values are rounded to align with the model grid resolution for consistency. 21 Figure 13: Ablation study of latent-space perturbations

  46. [54]

    The current position is used as the center point for the search

    Iterative Tracking: For each 6-hour time step: a. The current position is used as the center point for the search. b. If the current position is over water (determined by land-sea mask when enabled, however, we only enable this for tracking Dorian), the algorithm searches for ...

Pith tools

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