Pith. sign in

REVIEW 4 major objections 5 minor 8 cited by

FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read FourCastNet 3 claims a purely convolutional, one-step probabilistic model can match diffusion-based ensemble skill while keeping 60-day power spectra realistic.

desk verdict A serious contender with credible medium-range skill; the 60-day spectral evidence is internally inconsistent and must be fixed before the subseasonal claim ships. read the letter →

arxiv 2507.12144 v2 pith:7UQORXY5 submitted 2025-07-16 cs.LG physics.ao-ph

Boris Bonev , Thorsten Kurth , Ankur Mahesh , Mauro Bisson , Jean Kossaifi , Karthik Kashinath , Anima Anandkumar , William D. Collins
show 2 more authors
Michael S. Pritchard Alexander Keller
This is my paper
classification cs.LGphysics.ao-ph
keywords probabilisticweatherforecastingsphericalneuraloperatorensembleCRPSlossspectralfidelityhiddenMarkovmodelsubseasonalpredictionmachinelearning
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

FourCastNet 3 (FCN3) is a probabilistic machine-learning weather model built as a hidden Markov model: at each 6-hour step it maps the current atmospheric state plus a stochastic latent noise vector to the next state, so sampling the noise generates an ensemble. The paper's central claim is that this one-step scheme, trained end-to-end with a spatial and spectral CRPS objective, beats the standard physics-based ensemble on probabilistic skill, nearly matches the leading diffusion-based probabilistic model, and stays well calibrated with spread-skill ratios near one. Its distinctive claim is spectral fidelity: individual ensemble members keep the correct angular power-spectrum slopes for rollouts of 30 to 60 days, where competing probabilistic ML models show blurring or high-frequency buildup. Training on reanalysis data through 2016, a 50-member 15-day forecast costs about 60 seconds on one GPU, roughly 8 times faster than the diffusion baseline and 60 times faster than the conventional ensemble, which the authors argue makes large subseasonal ensembles practical.

What carries the argument

The load-bearing machinery is the combined objective and noise model. Stochasticity enters through a hidden Markov latent variable $z_n$ sampled from a spherical diffusion process parameterized in spectral space, with eight fixed length scales $k_T$, $\lambda=1$, and $\sigma=1$; this is meant to supply spatio-temporally correlated uncertainty. The loss is a composite of a spatially averaged ensemble CRPS and a spectral CRPS over spherical-harmonic coefficients $\hat{u}^m_\ell$, which the paper argues prevents the pointwise-CRPS failure mode in which shuffled but unrealistic members still score perfectly. The architecture carries the claim through spherical group convolutions: global filters parameterized by the convolution theorem on the sphere, $(\widehat{u\star k})^m_\ell = \hat{u}^m_\ell \hat{k}^0_\ell$, and local discrete-continuous spherical convolutions with Morlet-type wavelet filter bases.

What would settle it

Run the trained 50-member model on out-of-distribution initial conditions, for example the years after the training period or a cluster of heatwaves and cyclones, and compare each ensemble member's angular power spectral density to the reanalysis at 30- and 60-day leads. If the relative spectral error leaves the stated −0.2 to 0.2 band, or spread-skill ratios depart from one, the calibration and spectral-fidelity claims are falsified. A second, cheaper check: retrain with the spectral CRPS term removed; if the power-spectrum slopes remain correct, the paper's causal story about the loss is wrong.

Watch

Extended reading notes

Core claim

FCN3's discovery claim is that probabilistic weather forecasting does not need iterative diffusion sampling: a purely convolutional spherical neural operator, conditioned on noise from a spherical diffusion process, can produce calibrated ensembles in one forward pass per member. The architecture combines eight local discrete-continuous spherical convolutions for every one global spectral convolution, omits layer normalization, and predicts the next state directly rather than a residual tendency; the paper finds this suppresses high-frequency artifact buildup. The training objective couples a pointwise, spatially averaged CRPS with a spectral CRPS that scores the predictive distribution of each spherical-harmonic coefficient, weighted by mode multiplicity; the authors argue this is what keeps spatial correlations and power spectra physically realistic. Evaluated on 50-member ensembles over the out-of-sample year 2020, FCN3 outperforms the conventional physics ensemble and matches the leading diffusion model on 15 of 16 scoring channels at short lead times, at double the temporal resolution; spread-skill ratios approach one, rank histograms flatten, and the angular power spectral density matches the reanalysis ground truth with relative error between −0.2 and 0.2 at a 15-day lead.

Load-bearing premise

The paper's probabilistic claims rest on the assumption that eight hand-chosen spherical-diffusion noise scales, with fixed parameters $\lambda=1$ and $\sigma=1$, adequately represent the atmosphere's true correlated forecast uncertainty; if that noise is misspecified, the ensemble could appear calibrated under the CRPS loss without carrying physical uncertainty, and the 60-day spectral stability need not survive outside the reanalysis distribution.

Editorial extensions

If this is right

  • If the claims hold, operational centers could replace CPU-heavy ensemble weather prediction with one-GPU ML ensembles for medium-range and early-warning products.
  • A 60-day forecast at 0.25° and 6-hourly resolution in under 4 minutes on a single GPU makes subseasonal ensemble studies no longer compute-bound.
  • The spectral-stability claim implies the model can be autoregressed for 240 steps without the small-scale noise buildup that precedes blow-up in other probabilistic ML models.
  • The paper's success with a 50-member ensemble, larger than the 16 members used in training, indicates ensemble size can be scaled up without losing calibration.

Reading between the lines

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

  • Inference: the eight fixed noise scales are the fragile part; if the real forecast-error covariance is dominated by scales outside the chosen $k_T$ set, CRPS calibration could be achieved by construction rather than by physical representativeness. Adding data-assimilation perturbations to the noise would test this.
  • Inference: the speed comparison is at double temporal resolution, so the per-step cost advantage over the diffusion baseline is real but the wall-clock comparison partly reflects this resolution difference.
  • Inference: scoring each spectral mode's distribution instead of only pointwise marginals is a general principle that could transfer to any spatial probabilistic forecaster, including climate emulators and precipitation nowcasters.
  • Inference: the paper lists precipitation output and assimilation-uncertainty coupling as future work; those additions would test whether the spectral fidelity survives a moisture cycle and realistic initial-condition spread.
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

4 major / 5 minor

Summary. The manuscript presents FourCastNet 3 (FCN3), a probabilistic machine-learning weather forecasting system built from spherical convolutions, a hidden-Markov formulation with stochastic noise inputs, and a combined spatial-plus-spectral CRPS training objective. The authors report that FCN3 outperforms IFS-ENS, is competitive with GenCast, is well calibrated at 50 ensemble members, preserves angular power spectra up to 60-day lead times, runs 8–60× faster than the compared baselines, and can be trained at scale using a new domain-decomposition-based model/data parallelism scheme. Evaluation follows WeatherBench 2 practice on the out-of-sample year 2020, including CRPS, RMSE, ACC, spread-skill ratio, rank histograms, case studies, bias maps, and spectral diagnostics. The paper also documents the open-source Makani and torch-harmonics software.

Significance. If the claims hold, FCN3 is a significant advance: it demonstrates that a purely convolutional spherical architecture can match diffusion-based probabilistic skill at a fraction of the inference cost, and it reports spectral and rollout stability at subseasonal lead times that most ML weather models lack. The manuscript is commendably thorough in its evaluation: it uses a 50-member ensemble over the full year 2020, reports both aggregate scores and physical-fidelity diagnostics, includes case studies and bias maps, and makes training and inference code publicly available. The distributed-training contribution (spatial model parallelism with ensemble and batch parallelism) is also noteworthy and well documented. The main reservations concern the evidence for the headline 60-day spectral-stability claim and the interpretation of spectral fidelity given that the training loss directly optimizes it.

major comments (4)
  1. [Appendix F.7, Figure 23] The text and caption of Figure 23 contradict each other: the body text states that the figure shows 'four FCN3 forecast members initialized at 2018-01-01 00:00:00 UTC', while the caption says the angular PSD is 'averaged over daily initial conditions at 00:00:00 UTC and 12:00:00 UTC over the year 2020'. This is load-bearing because the abstract's claim of 'realistic spectra ... at extended lead times of up to 60 days' rests on the 1440-hour (60-day) panel of this figure. If the text is correct, the 60-day evidence is a single-initialization anecdote; if the caption is correct, the paper provides no sampling uncertainty for the 1440-hour PSD, unlike the 360-hour panel in Figure 5 which shows relative-error bounds. Please resolve the inconsistency and, in either case, provide a quantitative comparison with sampling uncertainty (e.g., relative-error bounds over initializations and ensemble members) at the 1440-hour lead time.
  2. [Section 5, Appendix E.1 (Eq. 51)] The spectral-fidelity result is partly by construction: the training objective includes a spectral CRPS term (Eq. 51) that directly encourages the spherical-harmonic coefficients of individual ensemble members to match the ground-truth distribution. The paper acknowledges this in Section 2 and Appendix E.1, and the acknowledgment is welcome. However, the abstract and Section 5 present spectral fidelity as an emergent advantage of the method ('retains perfectly the correct slopes'), which overstates the case for a quantity that is explicitly optimized. Please either temper these claims or provide an evaluation metric that is not a direct training target, such as spectral spread-skill or the spectral CRPS of a model ablated without the spectral loss term, to demonstrate that the 60-day spectral stability is not merely a re-statement of the loss function.
  3. [Abstract and Section 5 (speed comparisons)] The speed-up claims of '8x over GenCast' and '60x over IFS-ENS' compare inference on an NVIDIA H100 with GenCast on a Cloud TPU v5 and IFS on 96 AMD Epyc Rome CPUs, at different temporal resolutions. The paper acknowledges 'barring the differences in hardware and resolution' in Section 5, but the abstract states the speed-ups without qualification. Since speed is a central advertised advantage, please report at least one controlled comparison (same hardware, matched resolution) or clearly attach the hardware caveat to the abstract's speed figures.
  4. [Appendix B.7, Table 1, Figures 15–16] The ensemble's physical uncertainty representation depends on the hand-specified spherical diffusion noise with fixed parameters (eight length scales kT, lambda=1, sigma=1; Table 1, Eq. 27–28). The paper shows good aggregate calibration via spread-skill ratio and rank histograms, but it does not report scale-dependent or spatially resolved verification of ensemble spread, such as spread-skill as a function of spherical harmonic degree or lead-time-dependent spectral spread. Given that the 60-day spectral-stability claim concerns the behavior of individual members, please add a scale-resolved diagnostic (e.g., spectral spread-skill or banded rank histograms) to demonstrate that the ensemble represents uncertainty at the relevant wavelengths rather than only matching marginal pointwise distributions.
minor comments (5)
  1. [Figure 18 caption] The caption for Figure 18 (anomaly correlation coefficient, ACC) states 'Lower scores indicate better skill', but for ACC higher values indicate better skill; this is inconsistent with the caption of Figure 14. Please correct.
  2. [Section 3] In the paragraph describing the decoder, the sentence 'Both encoder and decoder encode do not perform any channel mixing' is grammatically broken; please rephrase, for example as 'Neither the encoder nor the decoder performs any channel mixing'.
  3. [Section G (Scaling FourCastNet 3)] At the end of the introductory paragraph, the sentence 'This is inspired from' is incomplete; please finish the sentence or delete the fragment.
  4. [Throughout] The model name is written inconsistently as 'FCN3' and 'FCN 3' (e.g., 'FCN 3 ensemble', 'FCN3' in captions); please use a single consistent abbreviation.
  5. [Appendix F.2] The text says FCN3 'outperforms IFS-ENS' on probabilistic and deterministic scores; Figure 12 and Figure 13 appear to support this for most channels and lead times, but the claim is made without a statistical significance test. A brief statement on the number of initializations (about 730 per year) and, if possible, confidence intervals for the score differences would strengthen the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: skill and calibration are out-of-sample against external baselines; spectral fidelity is an explicit training objective, and the 60-day claim is an evidentiary gap (Fig. 23 caption/text conflict), not a circular reduction.

full rationale

The derivation chain is self-contained: FCN3 is an autoregressive hidden Markov model with explicit stochastic forcing sampled from a published diffusion process (Appendix B.7, Eqs. 27-28, citing Palmer et al. [30]) and trained with a composite CRPS objective (Eqs. 48-51). The headline skill and calibration numbers are evaluated on out-of-sample 2020 initial conditions against independent external baselines (IFS-ENS and GenCast via WeatherBench 2), so they are not fitted inputs renamed as predictions. Self-citations to SFNO [10], Huge Ensembles [7,8], and DISCO [32] provide architectural lineage and context rather than load-bearing reductions, and no uniqueness theorem is imported from the authors' prior work to forbid alternatives. The one near-circular element is spectral fidelity: Eq. 51 computes CRPS per spherical-harmonic coefficient, and the reported angular PSD (Eq. 53) is the squared sum of those same coefficients, so matching the spectrum is explicitly optimized rather than emergent; the paper itself acknowledges this ('the combined CRPS loss function which enforces the correct local and global distribution'). However, the 60-day spectral claim goes beyond the 8-step autoregressive training horizon, so it is extrapolation rather than a by-construction result, and it is evaluated on held-out 2020 data. I also flag an evidence gap that is not circularity: Appendix F.7 text says Figure 23 shows 'four FCN3 forecast members initialized at 2018-01-01 00:00:00 UTC,' while the caption says 'averaged over daily initial conditions at 00:00:00 UTC and 12:00:00 UTC over the year 2020'; under either reading no sampling uncertainty is provided for the 1440-hour PSD, which weakens the abstract's 60-day spectral claim as a correctness matter, not as a circular one.

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

No new physical entities are introduced; the latent noise is a standard stochastic parameterization (Palmer et al. [30]). The free parameters are mostly loss weights and noise scales, some unreported (lambda_spectral), that the central spectral-fidelity and calibration claims depend on.

free parameters (4)
  • lambda_spectral = not reported
    Weight of the spectral CRPS term in Eq. (48); central to the spectral-fidelity claim, yet no value or tuning procedure is given in the text.
  • noise process parameters (kT, lambda, sigma) = kT in {3.08e-5,...,5.05e-1}, lambda=1, sigma=1
    Hand-selected in Table 1 / Eq. (28); define the distribution of the latent stochastic variable that drives the ensemble, so calibration depends on their adequacy.
  • local-to-global block ratio = 4 local : 1 global
    Chosen 'through experimentation' (Appendix C.5); affects skill and spectral behavior.
  • channel weights wc and temporal weights w_dt_c = defined in Table 4 and Eq. (49)
    Data-derived from ERA5 statistics; they shape the loss, but are not fitted to the target metric.
assumptions (6)
  • standard math Spherical harmonic transform and convolution theorems are accurate for the grids used.
    Used throughout the architecture (Appendix B).
  • domain assumption ERA5 reanalysis is a valid ground truth for the atmosphere and its spectrum.
    Training target and evaluation reference; spectral fidelity is defined relative to ERA5's T639 spectrum.
  • domain assumption The atmosphere can be modeled as a first-order Markov process on 6-hourly states with an additive latent noise.
    Hidden Markov formulation in Section 2 / Appendix A.2.
  • domain assumption Rotation-equivariant spherical convolutions are appropriate inductive biases for global weather.
    Section 3; assumes no preferred coordinate system beyond zonal/meridional anisotropy captured by kernels.
  • ad hoc to paper Direct state prediction (not tendency) is more stable in rollout; layer normalization should be omitted.
    Appendix C.5 and C.7; empirical, not theoretically justified.
  • ad hoc to paper Noise-centering (sign-flipping paired ensemble members) improves performance.
    Appendix E.3; unexplained and specific to this model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale." pith.science (2026). https://pith.science/paper/7UQORXY5

@misc{pith2026250712144,
  author       = {Pith},
  title        = {Pith review of: FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7UQORXY5}},
  note         = {Machine review of arXiv:2507.12144}
}
read the original abstract

FourCastNet 3 advances global weather modeling by implementing a scalable, geometric machine learning (ML) approach to probabilistic ensemble forecasting. The approach is designed to respect spherical geometry and to accurately model the spatially correlated probabilistic nature of the problem, resulting in stable spectra and realistic dynamics across multiple scales. FourCastNet 3 delivers forecasting accuracy that surpasses leading conventional ensemble models and rivals the best diffusion-based methods, while producing forecasts 8 to 60 times faster than these approaches. In contrast to other ML approaches, FourCastNet 3 demonstrates excellent probabilistic calibration and retains realistic spectra, even at extended lead times of up to 60 days. All of these advances are realized using a purely convolutional neural network architecture tailored for spherical geometry. Scalable and efficient large-scale training on 1024 GPUs and more is enabled by a novel training paradigm for combined model- and data-parallelism, inspired by domain decomposition methods in classical numerical models. Additionally, FourCastNet 3 enables rapid inference on a single GPU, producing a 60-day global forecast at 0.25{\deg}, 6-hourly resolution in under 4 minutes. Its computational efficiency, medium-range probabilistic skill, spectral fidelity, and rollout stability at subseasonal timescales make it a strong candidate for improving meteorological forecasting and early warning systems through large ensemble predictions.

Figures

Figures reproduced from arXiv: 2507.12144 by the authors.

Figure 1
Figure 1. Schematic of the FourCastNet 3 model. The model predicts the state of the atmosphere at the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of model- and data-parallelism for training of FourCastNet 3. In the given example, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Probabilistic skill of FourCastNet 3 relative to the ERA5 ground truth. Continuously ranked [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (23 more)
Figure 4
Figure 4. Figure 4: FourCastNet 3 prediction of storm Dennis initialized on 2020-02-11 at 00:00:00 UTC. The [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparison of angular power spectral densities of a single FourCastNet 3 ensemble member at [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Illustration of the equiangular (lat/lon-) grid in blue and a Gaussian grid in red. Both grids [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Depiction of random filters k(x) resulting from the filter parametrization. The spectral filter on the left is parameterized as a linear combination of globally supported spherical harmonics, whereas the local filter on the right is parameterized as a linear combinatio…
Figure 8
Figure 8. Figure 8: Illustration of the Morlet-type wavelet filter basis, used to represent localized filters on the [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: FourCastNet 3 encoder and decoder. Atmospheric, surface and auxiliary channels are all encoded [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]
Figure 10
Figure 10. Figure 10: Illustration of a single processor block. Green blocks indicate convolutions which take spatial [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Mean and variance of activations for 8 randomly initialized instances of FourCastNet 3 and [PITH_FULL_IMAGE:figures/full_fig_p025_11.png]
Figure 12
Figure 12. Figure 12: Continuously ranked probability scores (CRPS) averaged over initial conditions at 00:00:00 [PITH_FULL_IMAGE:figures/full_fig_p031_12.png]
Figure 13
Figure 13. Figure 13: Ensemble-mean root mean squared errors (RMSE), averaged over initial conditions at 00:00:00 [PITH_FULL_IMAGE:figures/full_fig_p031_13.png]
Figure 14
Figure 14. Figure 14: Ensemble-mean anomaly correlation coefficient (ACC), averaged over initial conditions at [PITH_FULL_IMAGE:figures/full_fig_p032_14.png]
Figure 15
Figure 15. Figure 15: Spread-skill ratios (SSR), averaged over initial conditions at 00:00:00 UTC and 12:00:00 UTC [PITH_FULL_IMAGE:figures/full_fig_p033_15.png]
Figure 16
Figure 16. Figure 16: Rank histograms of the ERA5 ground truth within the 50-member FourCastNet 3 ensemble at [PITH_FULL_IMAGE:figures/full_fig_p034_16.png]
Figure 17
Figure 17. Figure 17: Root mean squared errors (RMSE) of a single ensemble member, averaged over initial conditions [PITH_FULL_IMAGE:figures/full_fig_p035_17.png]
Figure 18
Figure 18. Figure 18: Single member anomaly correlation coefficient (ACC), averaged over initial conditions at [PITH_FULL_IMAGE:figures/full_fig_p036_18.png]
Figure 19
Figure 19. Figure 19: Depiction of a single 6-hour forecast initialized at 2018-01-01 00:00:00 UTC compared to the [PITH_FULL_IMAGE:figures/full_fig_p037_19.png]
Figure 20
Figure 20. Figure 20: Depiction of a single 6-hour forecast initialized at 2018-01-01 00:00:00 UTC compared to the [PITH_FULL_IMAGE:figures/full_fig_p038_20.png]
Figure 21
Figure 21. Figure 21: Depiction of a single 60-day forecast initialized at 2018-01-01 00:00:00 UTC compared to the [PITH_FULL_IMAGE:figures/full_fig_p039_21.png]
Figure 22
Figure 22. Figure 22: Bias of a single FourCastNet 3 ensemble member relative to the ERA5 ground truth averaged [PITH_FULL_IMAGE:figures/full_fig_p040_22.png]
Figure 23
Figure 23. Figure 23: Angular power spectral density of FCN3 forecasts at lead times of 6, 360 and 1440 hours, [PITH_FULL_IMAGE:figures/full_fig_p041_23.png]
Figure 24
Figure 24. Figure 24: Zonal power spectral density of ML weather forecasts at the 60 [PITH_FULL_IMAGE:figures/full_fig_p042_24.png]
Figure 25
Figure 25. Figure 25: Communicator hierarchy used in training and inference. To enable multiple forms of parallelism, [PITH_FULL_IMAGE:figures/full_fig_p042_25.png]
Figure 26
Figure 26. Figure 26: Illustration of domain decomposition within Makani for an azimuth group of 4 and a polar [PITH_FULL_IMAGE:figures/full_fig_p043_26.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 8 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. 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.

  3. Orca: Neural Operators for Causal Reasoning in Continuous Time

    cs.AI 2026-07 conditional novelty 6.0 of 10

    Orca extends structural causal models to continuous time with neural operators, enabling resolution-invariant dose-response and counterfactual trajectories on irregularly sampled cyclic systems.

  4. Physics constraints and response validation in discrete-time reduced-order modeling: from idealized turbulent systems to climate dynamics

    nlin.CD 2026-02 unverdicted novelty 6.0 of 10

    A framework builds stable neural models of turbulent dynamics by enforcing energy-preserving nonlinearities and causal constraints in discrete-time flow maps, demonstrated on Charney-DeVore and Lorenz-96 systems.

  5. MoWE : A Mixture of Weather Experts

    cs.LG 2025-09 conditional novelty 6.0 of 10

    MoWE, a ViT-based gating network, combines forecasts from Pangu, Aurora, and FCN3 with per-grid-point weights and beats each expert and the simple mean in RMSE.

  6. From Observation to Insight: Mechanistic World Models and the Quest for Autonomous Discovery

    cs.AI 2026-07 unverdicted novelty 5.0 of 10

    The paper proposes Mechanistic World Models — models organized as typed latent variables, a reusable mechanism library, and binding structures — as the route from AI forecasting to autonomous discovery.

  7. CRPS-LAM: Probabilistic Regional Weather Forecasting with Continuous Ranked Probability Score

    cs.LG 2025-10 conditional novelty 5.0 of 10

    CRPS-LAM produces 57-hour probabilistic limited-area forecasts on MEPS at diffusion-comparable accuracy with single-forward-pass sampling, roughly 39x faster than Diffusion-LAM.

  8. The Rise of AI in Weather and Climate Information and its Impact on Global Inequality

    physics.ao-ph 2026-03 conditional novelty 4.0 of 10

    AI weather and climate tools inherit Northern-controlled data and compute, risking worse forecasts and maladaptation for the Global South rather than democratizing climate information.

Reference graph

Works this paper leans on

77 extracted references · 20 canonical work pages · cited by 8 Pith papers

  1. [1]

    The origins of computer weather prediction and climate modeling

    Peter Lynch. The origins of computer weather prediction and climate modeling. Journal of Computational Physics , 227:3431–3444, 2008. ISSN 0021-9991. doi:https://doi.org/10.1016/j.jcp.2007.02.034. URL https://www.sciencedirect.com/science/ article/pii/S0021999107000952. Predicting weather, climate and extreme events

    doi:10.1016/j.jcp.2007.02.034 2008

  2. [2]

    FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators, 2 2022

    Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, Pedram Hassanzadeh, Karthik Kashinath, and Animashree Anandkumar. FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators, 2 2022. URL ht...

    arXiv 2022

  3. [3]

    Pangu-Weather: A 3D High-Resolution Model for Fast and Accurate Global Weather Forecast, 11 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, 11 2022. URL http://arxiv.org/abs/2211.02556

    arXiv 2022

  4. [4]

    GraphCast: Learning skillful medium-range global weather forecasting, 12 2022

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Alexander Pritzel, Suman Ravuri, Timo Ewalds, Ferran Alet, Zach Eaton-Rosen, Weihua Hu, Alexander Merose, Stephan Hoyer, George Holland, Jacklynn Stott, et al. GraphCast: Learning skillful medium-range global weather forecasting, 12 2022. URL http://arxiv.org/abs/2212.12794

    arXiv 2022

  5. [5]

    FourCastNet: Accelerating Global High-Resolution Weather Forecasting using Adaptive Fourier Neural Operators, 2022

    Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Animashree Anandkumar. FourCastNet: Accelerating Global High-Resolution Weather Forecasting using Adaptive Fourier Neural Operators, 2022. URL https://arxiv.org/abs/2208.05419

    arXiv 2022

  6. [6]

    Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, and Matthew Willson

    Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R. Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, and Matthew Willson. Probabilistic weather forecasting with machine learning. Nature, 1 2024. ISSN 14764687. doi:10.1038/s41586-024-08252-9

    doi:10.1038/s41586-024-08252-9 2024

  7. [7]

    Huge Ensembles Part I: Design of Ensemble Weather Forecasts using Spherical Fourier Neural Operators, 8 2024

    Ankur Mahesh, William Collins, Boris Bonev, Noah Brenowitz, Yair Cohen, Joshua Elms, Peter Harrington, Karthik Kashinath, Thorsten Kurth, Joshua North, Travis OBrien, Michael Pritchard, David Pruitt, Mark Risser, Shashank Subramanian, and Jared Willard. Huge Ensembles Part I: Design of Ensemble Weather Forecasts using Spherical Fourier Neural Operators, 8...

    arXiv 2024

  8. [8]

    Huge Ensembles Part II: Properties of a Huge Ensemble of Hindcasts Generated with Spherical Fourier Neural Operators, 8 2024

    Ankur Mahesh, William Collins, Boris Bonev, Noah Brenowitz, Yair Cohen, Peter Harrington, Karthik Kashinath, Thorsten Kurth, Joshua North, Travis OBrien, Michael Pritchard, David Pruitt, Mark Risser, Shashank Subramanian, and Jared Willard. Huge Ensembles Part II: Properties of a Huge Ensemble of Hindcasts Generated with Spherical Fourier Neural Operators...

    arXiv 2024

Show all 77 references
  1. [9]

    Weyn, Dale R

    Jonathan A. Weyn, Dale R. Durran, Rich Caruana, and Nathaniel Cresswell-Clay. Sub-seasonal forecasting with a large ensemble of deep-learning weather prediction models, 2 2021. URL http: //arxiv.org/abs/2102.05107http://dx.doi.org/10.1029/2021MS002502

    2021 arXiv

  2. [10]

    Spherical Fourier Neural Operators: Learning Stable Dynamics on the Sphere

    Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar. Spherical Fourier Neural Operators: Learning Stable Dynamics on the Sphere. Proceedings of the 40th International Conference on Machine Learning , 202:2806–...

    2023 arXiv

  3. [11]

    Clark, Brian Henn, James Duncan, Noah D

    Oliver Watt-Meyer, Gideon Dresdner, Jeremy McGibbon, Spencer K. Clark, Brian Henn, James Duncan, Noah D. Brenowitz, Karthik Kashinath, Michael S. Pritchard, Boris Bonev, Matthew E. Peters, and Christopher S. Bretherton. ACE: A fast, skillful learned global atmospheric model fo...

    2023 arXiv

  4. [12]

    Durran, Raul A

    Matthias Karlbauer, Nathaniel Cresswell-Clay, Dale R. Durran, Raul A. Moreno, Thorsten Kurth, Boris Bonev, Noah Brenowitz, and Martin V. Butz. Advancing Parsimonious Deep Learning Weather 9 FourCastNet 3 Prediction Using the HEALPix Mesh. Journal of Advances in Modeling Earth ...

    2024 doi

  5. [13]

    Lucie: A lightweight uncoupled climate emulator with long-term stability and physical consistency for o(1000)-member ensembles, 4 2025

    Haiwen Guan, Troy Arcomano, Ashesh Chattopadhyay, and Romit Maulik. Lucie: A lightweight uncoupled climate emulator with long-term stability and physical consistency for o(1000)-member ensembles, 4 2025. URL http://arxiv.org/abs/2405.16297

    2025 arXiv

  6. [14]

    Espinosa, Raul Moreno, and Matthias Karlbauer

    Nathaniel Cresswell-Clay, Bowen Liu, Dale Durran, Zihui Liu, Zachary I. Espinosa, Raul Moreno, and Matthias Karlbauer. A deep learning earth system model for efficient simulation of the observed climate, 2 2025. URL http://arxiv.org/abs/2409.16247

    2025

  7. [15]

    Dynamical errors in machine learning forecasts, 4 2025

    Zhou Fang and Gianmarco Mengaldo. Dynamical errors in machine learning forecasts, 4 2025. URL http://arxiv.org/abs/2504.11074

    2025 arXiv

  8. [16]

    Brenner, and Stephan Hoyer

    Dmitrii Kochkov, Janni Yuval, Ian Langmore, Peter Norgaard, Jamie Smith, Griffin Mooers, Milan Kl¨ ower, James Lottes, Stephan Rasp, Peter D¨ uben, Sam Hatfield, Peter Battaglia, Alvaro Sanchez-Gonzalez, Matthew Willson, Michael P. Brenner, and Stephan Hoyer. Neural General Ci...

    2023 doi

  9. [17]

    WeatherBench 2: A benchmark for the next generation of data-driven global weather models, 8 2023

    Stephan Rasp, Stephan Hoyer, Alexander Merose, Ian Langmore, Peter Battaglia, Tyler Russel, Alvaro Sanchez-Gonzalez, Vivian Yang, Rob Carver, Shreya Agrawal, Matthew Chantry, Zied Ben Bouallegue, Peter Dueben, Carla Bromberg, Jared Sisk, et al. WeatherBench 2: A benchmark for ...

    2023 arXiv

  10. [18]

    Brenowitz, Yair Cohen, Jaideep Pathak, Ankur Mahesh, Boris Bonev, Thorsten Kurth, Dale R

    Noah D. Brenowitz, Yair Cohen, Jaideep Pathak, Ankur Mahesh, Boris Bonev, Thorsten Kurth, Dale R. Durran, Peter Harrington, and Michael S. Pritchard. A Practical Probabilistic Benchmark for AI Weather Models. 1 2024. URL http://arxiv.org/abs/2401.15305

    2024 arXiv

  11. [19]

    Fixing the Double Penalty in Data-Driven Weather Forecasting Through a Modified Spherical Harmonic Loss Function, 1 2025

    Christopher Subich, Syed Zahid Husain, Leo Separovic, and Jing Yang. Fixing the Double Penalty in Data-Driven Weather Forecasting Through a Modified Spherical Harmonic Loss Function, 1 2025. URL http://arxiv.org/abs/2501.19374

    2025 arXiv

  12. [20]

    AtmoRep: A stochastic model of atmosphere dynamics using large scale representation learning, 8

    Christian Lessig, Ilaria Luise, Bing Gong, Michael Langguth, Scarlet Stadtler, and Martin Schultz. AtmoRep: A stochastic model of atmosphere dynamics using large scale representation learning, 8

  13. [21]

    Continuous Ensemble Weather Forecasting with Diffusion models, 10 2024

    Martin Andrae, Tomas Landelius, Joel Oskarsson, and Fredrik Lindsten. Continuous Ensemble Weather Forecasting with Diffusion models, 10 2024. URL http://arxiv.org/abs/2410.05431

    2024 arXiv

  14. [22]

    Simon Lang, Mihai Alexe, Mariana C. A. Clare, Christopher Roberts, Rilwan Adewoyin, Zied Ben Bouall` egue, Matthew Chantry, Jesper Dramsch, Peter D. Dueben, Sara Hahner, Pedro Maciel, Ana Prieto-Nemesio, Cathal O’Brien, Florian Pinault, Jan Polster, et al. AIFS-CRPS: Ensemble ...

    2024 arXiv

  15. [23]

    Denoising Diffusion Probabilistic Models, 6 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models, 6 2020. URL http://arxiv.org/abs/2006.11239

    2020 arXiv

  16. [24]

    Strictly Proper Scoring Rules, Prediction, and Estimation

    Tilmann Gneiting and Adrian E Raftery. Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association , 102:359–378, 3 2007. ISSN 0162-1459. doi:10.1198/016214506000001437. URL http://www.tandfonline.com/doi/abs/10. 1198/016214506000001437

    2007 doi

  17. [25]

    Lecture notes in computational science and engineering ; 80

    Peter Lauritzen, Christiane Jablonowski, Mark Taylor, and Ramachandran Nair.Numerical techniques for global atmospheric models . Lecture notes in computational science and engineering ; 80. Springer, Berlin ;, 2011. ISBN 9783642116391

    2011

  18. [26]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. Advances in Neural Information Processing Systems, 6 2017. URL http://arxiv.org/abs/1706.03762

    2017 arXiv

  19. [27]

    Battaglia

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning mesh-based simulation with graph networks. 6 2021. URL http://arxiv.org/abs/2010.03409

    2021 arXiv

  20. [28]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. 1

  21. [29]

    Evaluation of ecmwf forecasts

    Thomas Haiden, Martin Janousek, Fr´ ed´ eric Vitart, Maliko Tanguy, Fernando Prates, and Matthieu Chevalier. Evaluation of ecmwf forecasts. ECMWF Newsletter , 2024. doi:10.21957/52F2F31351. URL https://www.ecmwf.int/en/elibrary/81582-evaluation-ecmwf-forecasts

    2024 doi

  22. [30]

    Stochastic Parametrization and Model Uncertainty

    T N Palmer, R Buizza, F Doblas-Reyes, T Jung, M Leutbecher, G J Shutts, M Steinheimer, and A Weisheimer. Stochastic Parametrization and Model Uncertainty. Technical report, ECMWF, 2009. URL http://www.ecmwf.int/publications/. 10 FourCastNet 3

    2009

  23. [31]

    Price, and Jason D

    Jeremy Ocampo, Matthew A. Price, and Jason D. McEwen. Scalable and Equivariant Spherical CNNs by Discrete-Continuous (DISCO) Convolutions, 9 2022. URL http://arxiv.org/abs/2209.13603

    2022 arXiv

  24. [32]

    Neural Operators with Localized Integral and Differential Kernels, 2 2024

    Miguel Liu-Schiaffini, Julius Berner, Boris Bonev, Thorsten Kurth, Kamyar Azizzadenesheli, and Anima Anandkumar. Neural Operators with Localized Integral and Differential Kernels, 2 2024. URL http://arxiv.org/abs/2402.16845

    2024 arXiv

  25. [33]

    A ConvNet for the 2020s, 1 2022

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s, 1 2022. URL http://arxiv.org/abs/2201.03545

    2022 arXiv

  26. [34]

    Gaussian Error Linear Units (GELUs)

    Dan Hendrycks and Kevin Gimpel. Gaussian Error Linear Units (GELUs). 6 2016. URL http: //arxiv.org/abs/1606.08415

    2016 arXiv

  27. [35]

    Hechtman, and Shibo Wang

    Yuanzhong Xu, HyoukJoong Lee, Dehao Chen, Hongjun Choi, Blake A. Hechtman, and Shibo Wang. Automatic Cross-Replica Sharding of Weight Update in Data-Parallel Training. CoRR, abs/2004.13336, 2020. URL https://arxiv.org/abs/2004.13336

    2004 arXiv

  28. [36]

    DeepSpeed: System Optimizations Enable Training Deep Learning Models with Over 100 Billion Parameters

    Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. DeepSpeed: System Optimizations Enable Training Deep Learning Models with Over 100 Billion Parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’20...

    2020

  29. [37]

    ZeRO: Memory Optimizations Toward Training Trillion Parameter Models

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models. ArXiv, May 2020. URL https://www.microsoft.com/en-us/research/publication/ zero-memory-optimizations-toward-training-trillion-parameter-models/

    2020

  30. [38]

    GPT-4 Technical Report, 2024

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, et al. GPT-4 Technical Report, 2024. URL https: //ar...

    2024 arXiv

  31. [39]

    The Llama 3 Herd of Models, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, et al. The Llama 3 Herd of Models, 2024. URL https:...

    2024 arXiv

  32. [40]

    The ERA5 global reanalysis

    Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, Andr´ as Hor´ anyi, Joaqu´ ın Mu˜ noz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, Adrian Simmons, Cornel Soci, Saleh Abdalla, Xavier Abellan, Gianpaolo Balsamo, et al. The ERA5 global reanalys...

    1999 doi

  33. [41]

    Data-driven ensemble forecasting with the aifs

    Mihai Alexe, Simon Lang, Mariana Clare, Martin, Leutbecher, Christopher Roberts, Linus Magnusson, Matthew Chantry, Rilwan Adewoyin, Ana Prieto-Nemesio, Jesper Dramsch, Florian Pinault, and Baudouin Raoult. Data-driven ensemble forecasting with the aifs. https://www.ecmwf.int/e...

  34. [42]

    Consistency Models, 3 2023

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency Models, 3 2023. URL http://arxiv.org/abs/2303.01469

    2023 arXiv

  35. [43]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks, 6 2014. URL http://arxiv. org/abs/1406.2661

    2014 arXiv

  36. [44]

    Driscoll and D.M

    J.R. Driscoll and D.M. Healy. Computing Fourier Transforms and Convolutions on the 2-Sphere. Advances in Applied Mathematics , 15:202–250, 6 1994. ISSN 01968858. doi:10.1006/aama.1994.1008. URL https://linkinghub.elsevier.com/retrieve/pii/S0196885884710086

    1994

  37. [45]

    Jason. D. McEwen and Yves Wiaux. A novel sampling theorem on the sphere, 10 2011. URL http://arxiv.org/abs/1110.6298http://dx.doi.org/10.1109/TSP.2011.2166394

    2011 arXiv

  38. [46]

    A method for numerical integration on an automatic computer

    C W Clenshaw and A R Curtis. A method for numerical integration on an automatic computer. Numerische Mathematik , 2:197–205, 1960. ISSN 0945-3245. doi:10.1007/BF01386223. URL https: //doi.org/10.1007/BF01386223

    1960 doi

  39. [47]

    Golub and John H

    Gene H. Golub and John H. Welsch. Calculation of Gauss quadrature rules. Mathematics of Computation, 23:221–230, 1969. ISSN 0025-5718. doi:10.1090/S0025-5718-69-99647-1. URL https: //www.ams.org/mcom/1969-23-106/S0025-5718-69-99647-1/

    1969 doi

  40. [48]

    Cohen and Max Welling

    Taco S. Cohen and Max Welling. Group Equivariant Convolutional Networks, 2 2016. URL http://arxiv.org/abs/1602.07576. 11 FourCastNet 3

    2016 arXiv

  41. [49]

    Efficient spherical harmonic transforms aimed at pseudospectral numerical simulations

    Nathana¨ el Schaeffer. Efficient spherical harmonic transforms aimed at pseudospectral numerical simulations. Geochemistry, Geophysics, Geosystems , 14:751–758, 3 2013. ISSN 15252027. doi:10.1002/ggge.20071

    2013 doi

  42. [50]

    Cobb, Christopher G

    Oliver J. Cobb, Christopher G. R. Wallis, Augustine N. Mavor-Parker, Augustin Marignier, Matthew A. Price, Mayeul d’Avezac, and Jason D. McEwen. Efficient Generalized Spherical CNNs, 10 2020. URL http://arxiv.org/abs/2010.11661

    2020 arXiv

  43. [51]

    Giraldo, J.S

    F.X. Giraldo, J.S. Hesthaven, and T. Warburton. Nodal High-Order Discontinuous Galerkin Methods for the Spherical Shallow Water Equations. Journal of Computational Physics , 181:499–525, 9 2002. ISSN 00219991. doi:10.1006/jcph.2002.7139. URL https://linkinghub.elsevier.com/ret...

    2002

  44. [52]

    Randall J. LeVeque. Numerical Methods for Conservation Laws . Birkh¨ auser Basel, 1992. ISBN 978-3-7643-2723-1. doi:10.1007/978-3-0348-8629-1. URL http://link.springer.com/10.1007/ 978-3-0348-8629-1

    1992 doi

  45. [53]

    U-Net: Convolutional Networks for Biomedical Image Segmentation, 5 2015

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation, 5 2015. URL http://arxiv.org/abs/1505.04597

    2015 arXiv

  46. [54]

    F.J. Harris. On the use of windows for harmonic analysis with the discrete Fourier transform. Proceedings of the IEEE , 66:51–83, 1978. ISSN 0018-9219. doi:10.1109/PROC.1978.10837. URL http://ieeexplore.ieee.org/document/1455106/

    1978

  47. [55]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library, 12 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K¨ opf, Edward Yang, Zach DeVito, Martin Raison, et al. PyTorch: An Imperative Style, High-Performance Dee...

    2019 arXiv

  48. [56]

    Martin, Kenneth L

    Yingzhou Li, Haizhao Yang, Eileen R. Martin, Kenneth L. Ho, and Lexing Ying. Butterfly Factorization. Multiscale Modeling & Simulation , 13:714–732, 1 2015. ISSN 1540-3459. doi:10.1137/15M1007173. URL http://epubs.siam.org/doi/10.1137/15M1007173

    2015 doi

  49. [57]

    Accurate medium-range global weather forecasting with 3D neural networks

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619:533–538, 7 2023. ISSN 14764687. doi:10.1038/s41586-023-06185-3

    2023 doi

  50. [58]

    L. Chen, X. Zhong, F. Zhang, Y. Xu, Y. Chen, F. Zhu, H. Li, Y. Qian, and L. Chen. FuXi: a cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6:190, 2023. doi:10.1038/s41612-023-00512-1

    2023 doi

  51. [59]

    Going deeper with Image Transformers

    Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Herv´ e J´ egou. Going deeper with Image Transformers. CoRR, abs/2103.17239, 2021. URL https://arxiv.org/abs/ 2103.17239

    2021 arXiv

  52. [60]

    Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification, 2 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification, 2 2015. URL http://arxiv.org/abs/1502. 01852

    2015

  53. [61]

    Exploring the design space of deep-learning-based weather forecasting systems, 10 2024

    Shoaib Ahmed Siddiqui, Jean Kossaifi, Boris Bonev, Christopher Choy, Jan Kautz, David Krueger, and Kamyar Azizzadenesheli. Exploring the design space of deep-learning-based weather forecasting systems, 10 2024. URL http://arxiv.org/abs/2410.07472

    2024 arXiv

  54. [62]

    Alias-Free Generative Adversarial Networks, 6 2021

    Tero Karras, Miika Aittala, Samuli Laine, Erik H¨ ark¨ onen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-Free Generative Adversarial Networks, 6 2021. URL http://arxiv.org/abs/2106.12423

    2021 arXiv

  55. [63]

    Abaza, F

    Vincent Fortin, M. Abaza, F. Anctil, and R. Turcotte. Why should ensemble spread match the RMSE of the ensemble mean? Journal of Hydrometeorology, 15:1708–1713, 8 2014. ISSN 15257541. doi:10.1175/JHM-D-14-0008.1

    2014 doi

  56. [64]

    Strictly Proper Scoring Rules, Prediction, and Estimation,

    Tilmann Gneiting and Adrian E Raftery. Strictly Proper Scoring Rules, Prediction, and Estimation,

  57. [65]

    Estimation of the Continuous Ranked Probability Score with Limited Information and Applications to Ensemble Weather Forecasts

    Micha¨ el Zamo and Philippe Naveau. Estimation of the Continuous Ranked Probability Score with Limited Information and Applications to Ensemble Weather Forecasts. Mathematical Geosciences, 50:209–234, 2 2018. ISSN 18748953. doi:10.1007/s11004-017-9709-7

    2018 doi

  58. [66]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization, 12 2014. URL http://arxiv.org/abs/1412.6980

    2014 arXiv

  59. [67]

    Data Assimilation in the Geosciences - An overview on methods, issues and perspectives, 9 2017

    Alberto Carrassi, Marc Bocquet, Laurent Bertino, and Geir Evensen. Data Assimilation in the Geosciences - An overview on methods, issues and perspectives, 9 2017. URL http://arxiv.org/ abs/1709.02798

    2017 arXiv

  60. [68]

    Neural Networks for Postprocessing Ensemble Weather Forecasts

    Stephan Rasp and Sebastian Lerch. Neural Networks for Postprocessing Ensemble Weather Forecasts. Monthly Weather Review , pages 3885–3900, 2018. doi:10.1175/MWR-D-18. URL https://doi.org/ 10.1175/MWR-D-18- . 12 FourCastNet 3

    2018 doi

  61. [69]

    Hunt, and Edward Ott

    Troy Arcomano, Istvan Szunyogh, Jaideep Pathak, Alexander Wikner, Brian R. Hunt, and Edward Ott. A Machine Learning-Based Global Atmospheric Forecast Model. Geophysical Research Letters, 47, 5 2020. ISSN 0094-8276. doi:10.1029/2020GL087776. URL https://agupubs.onlinelibrary. w...

    2020 doi

  62. [70]

    Thomas M. Hamill. Interpretation of Rank Histograms for Verifying Ensemble Forecasts. Monthly Weather Review , 129:550–560, 3 2001. ISSN 0027-0644. doi:10.1175/1520-0493(2001)129<0550:IORHFV>2.0.CO;2. URL http://journals.ametsoc. org/doi/10.1175/1520-0493(2001)129<0550:IORHFV>2.0.CO;2

    2001 doi

  63. [71]

    Tulloch and K

    R. Tulloch and K. S. Smith. A theory for the atmospheric energy spectrum: Depth-limited temperature anomalies at the tropopause. Proceedings of the National Academy of Sciences , 103(40): 14690–14694, 2006. doi:10.1073/pnas.0605494103. URL https://www.pnas.org/doi/abs/10.1073/...

    2006 doi

  64. [72]

    Pytorch fsdp: Experiences on scaling fully sharded data parallel

    Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, Alban Desmaison, Can Balioglu, Pritam Damania, Bernard Nguyen, Geeta Chauhan, et al. Pytorch fsdp: Experiences on scaling fully sharded data paralle...

    2023

  65. [76]

    6Further detail on the simultaneous model- and data-parallelism are provided in Appendix G

    This training stage was carried out on 512 NVIDIA A100 GPUs on the National Energy Research Scientific Computing Center’s Perlmutter Supercomputer and took 15 hours to complete. 6Further detail on the simultaneous model- and data-parallelism are provided in Appendix G. 28 Four...

    1980

  66. [2019]

    A notable exception is 2 meter temperature t2m, where we observe better performance with GenCast

    On 15 out of the 16 channels in the scoring protocol, FCN3 matches and even slightly outperforms Gencast’s forecasts at short lead times. A notable exception is 2 meter temperature t2m, where we observe better performance with GenCast. The slightly better rollout skill of GenC...

    2020

  67. [2020]

    URL http://arxiv.org/abs/2001.08361

    2001 arXiv

  68. [2023]

    URL http://arxiv.org/abs/2308.13280

    arXiv

  69. [2024]

    Accessed: 2025-07-06

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.