Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

By learning per-grid-point, lead-time-dependent weights with a lightweight transformer, the authors combine three existing AI weather models into a single deterministic forecast that is more accurate than any of them individually, cutting R

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

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.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A sensible idea with thin evidence: the dynamic gating claims need a static-weight baseline and more than one test year before I trust the 10% figure. the 3 major comments →

arxiv 2509.09052 v1 pith:B2MQEUZB submitted 2025-09-10 cs.LG cs.AIphysics.ao-phphysics.geo-ph

MoWE : A Mixture of Weather Experts

Dibyajyoti Chakraborty , Romit Maulik , Peter Harrington , Dallas Foster , Mohammad Amin Nabian , Sanjay Choudhry This is my paper
classification cs.LG cs.AIphysics.ao-phphysics.geo-ph
keywords mixture of expertsweather forecastinggating networkvision transformermodel ensemblingdeterministic forecastRMSEdata-driven weather models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper argues that the plateau in data-driven weather forecasting can be broken without designing a new architecture. Instead, the authors propose MoWE, a lightweight gating network that learns to blend the forecasts of several existing expert models at every grid point, conditioned on forecast lead time. Their experiments on three state-of-the-art models show the blended forecast outperforms every expert and also beats the simple average of experts, achieving up to 10% lower RMSE than the best individual model at two-day lead times. The result matters because it offers a computationally cheap and scalable path to better forecasts: rather than competing, models can be combined to exploit their complementary strengths.

Core claim

The central discovery is that a synthesized deterministic forecast formed by a learned, spatiotemporally varying weighted combination of expert forecasts—where the weights are produced by a Vision Transformer-based gating network conditioned on lead time—achieves lower RMSE than any individual expert and than a simple mean. In tests across nine atmospheric variables and lead times from 6 to 48 hours, MoWE consistently outperforms both, with the largest gains (about 10%) at 48 hours. The learned weight maps are not uniform: they favor the strongest expert at short leads, spread more evenly at longer leads, and show geography-dependent structure, indicating the network is exploiting complement

What carries the argument

The central object is the gating network fgate, a Vision Transformer (specifically a DiT-style block) that takes the stacked forecast fields from all experts plus a lead-time embedding and outputs per-expert weight maps Wi and a bias map b. The final prediction is a per-grid-point softmax-normalized weighted sum of the experts' forecasts plus the bias. This machinery turns the model combination problem into a supervised learning problem: the gating network is trained end-to-end by minimizing MSE against ground truth, making the blend adaptive in space and over lead time.

Load-bearing premise

The load-bearing premise is that the improvement over simple averaging comes from the learned dynamic gating, yet the paper does not compare against a static, per-grid-point weighting fit on the same training data; if such a fixed weighting matched MoWE's test error, the dynamic conditioning would not be necessary.

What would settle it

Train a 'static MoWE' in which the gating network is replaced by a set of learned per-grid-point weights and a bias that do not depend on lead time, fit to the same training data. If its test RMSE on 2015 is statistically indistinguishable from MoWE's, then the dynamic, lead-time-conditioned gating is not the source of the reported 10% improvement.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If MoWE's gains hold across a broader set of experts, forecast skill can be improved by combining existing models rather than training new ones.
  • The gating network is light (25M parameters) and cheap to train compared with the experts, suggesting model combination can be a cost-effective path to better forecasts.
  • Because the weight maps are interpretable, they could be used to diagnose which model is most skilful where and at what lead time, informing future model development.
  • The framework scales in principle: as more expert models become available, they can be added without retraining from scratch, just extending the input channels and weight maps.
  • The approach can be viewed as a learned bias correction and model selection mechanism, applicable beyond the three experts tested.

Where Pith is reading between the lines

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

  • A natural next test is whether the learnt weights generalise to lead times beyond 48 hours or to variables not present in the training data; if they do, the gating network could replace per-lead-time retraining.
  • The absence of an ablation against a static per-grid-point weighting means the paper does not yet isolate whether dynamic, lead-time-conditioned gating is the source of the gains; a fixed weighting trained on the same data would settle this.
  • The authors' framing suggests a broader principle: in any domain where multiple learned predictors are available and no one dominates everywhere, a learned mixture may outperform even the best single model—weather is a testbed for this idea.
  • With an online training setup and compression of expert channels, the method could be extended to more experts and higher resolutions, potentially making multi-model fusion a standard practice in operational forecasting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes MoWE, a mixture-of-experts framework for deterministic global weather forecasting. Instead of training a new standalone model, MoWE uses a lightweight Vision Transformer-based gating network to combine the outputs of three pretrained AI weather models (Pangu, Aurora, and FCN3). The network produces per-grid-point weight maps and a bias field, conditioned on forecast lead time, and is trained on ERA5-initialized expert forecasts from 1980 to 2014 with an MSE loss. The authors evaluate on 2015 and report RMSE improvements over every individual expert and over a simple mean, including a claimed up-to-10% reduction relative to the best expert at a 2-day lead. A small capacity ablation compares Base (25M) and Small (9M) gating networks.

Significance. If the claimed improvements are robust, this is a practically useful and computationally cheap way to push deterministic forecast skill by exploiting complementarity among existing AI weather models. The paper is clearly written, uses a standard train/test split, and the authors state that the code is open-sourced, which are strengths. However, the empirical evidence is currently thin: the central novelty is dynamic, lead-time-conditioned gating, yet no learned static-weight baseline is tested, and all conclusions rest on one test year without uncertainty quantification. The idea is significant for the AI-weather community, but the paper needs additional experiments before the central claim can be accepted.

major comments (3)
  1. [Section 3 / Eq. (1)] The central claim that dynamic, lead-time-conditioned gating is responsible for the gains is not tested against a learned static weighting. Eq. (1) defines W_i as a full weight map, but W_i could instead be a constant per grid point (or per variable) fit on the same 1980-2014 training data by minimizing the same MSE. Such a static baseline could also exploit spatial model strengths, and the paper does not show that MoWE's dynamic weight maps improve over it. This ablation is load-bearing because the Conclusion attributes the improvement to isolating expert advantages 'to specific locations and lead times.' Please add a static per-grid-point or per-variable weighting baseline and report its test RMSE alongside MoWE.
  2. [Section 3, Figure 2] All results are for a single test year (2015) and are presented as RMSE curves without confidence intervals or significance tests. The claimed 10% improvement over the best expert, and the stated superiority over the simple mean, could be within sampling variability for some variables. Please provide per-variable RMSE tables with bootstrapped uncertainty estimates, or evaluate over multiple years (e.g., 2016-2020), and state whether the improvements are consistent across all start dates in 2015. Without this, the strength of the empirical claim is difficult to assess.
  3. [Abstract / Section 3] The manuscript never states which variable and lead time attain the claimed 10% improvement. Figure 2 shows qualitative curves, and Table 2 reports only the Base-vs-Small ablation, not the MoWE-vs-expert comparison. The headline result is therefore not reproducible from the text. Please add a table with exact RMSE values by variable and lead time (e.g., 6, 12, 24, 48 h) for MoWE, each expert, and the simple mean, and explicitly identify the variable and lead time corresponding to the 10% figure.
minor comments (5)
  1. [Section 2] The description of FCN3 as using 'a hidden Markov model' appears inaccurate: FCN3 is a spherical-neural-operator-based probabilistic model, not an HMM. Please correct or clarify this sentence.
  2. [References] Reference [28] is cited as 'Jane Doe, Piggycast: Improving weather prediction' with a Substack URL. This looks like an informal placeholder rather than a citable source. Please replace it with a proper reference to prior multi-model combination work, or remove it if it is not relevant.
  3. [Table 2 / Figure 2] Figure 2's caption says 'nine different atmospheric variables,' but Table 2 lists eleven entries (t2m, t500, t850, u10m, u500, u850, v10m, v500, v850, z500, z850). Please reconcile the variable count.
  4. [Abstract / Conclusion] The abstract says 'up to a 10% lower RMSE' while the Conclusion says 'a significant RMSE reduction of over 10%.' Please make these claims consistent and tie them to the same variable/lead time.
  5. [Section 3] The claim of 'significantly lower computational resources' is not quantified. Please report training time, GPU count, and total compute for the gating network, and ideally compare it to the cost of the expert models.

Circularity Check

0 steps flagged

No significant circularity: MoWE is a held-out supervised fusion model, not a derivation that reduces to its inputs.

full rationale

The claimed chain is empirical rather than definitional. MoWE uses Eq. (1), Ŷ = Σ W_i ⊙ E_i + b, and learns (W,b) = fgate(E1,...,EN,t,z) by minimizing MSE against ERA5 on 1980–2014, then evaluates on 2015. The simple-mean baseline is the special case W_i = 1/N, b = 0, so on the training set MoWE's optimization can only match or beat the mean; however, every reported RMSE in Table 2 and Figure 2 is for held-out 2015, so the advantage over the mean is a generalization result, not forced by construction. No equation reduces to a fitted constant renamed as prediction; no uniqueness theorem is invoked; and the DiT/ViT design is attributed to external prior work ([29], [30]) without making the central claim depend on an author-overlapping citation. The deepest legitimate concern is experimental: the paper does not ablate a static per-grid-point weighting fit on the same training data, so the conclusion that dynamic lead-time-conditioned gating is the source of the improvement is under-supported. That is a control gap and a correctness risk, not circularity. The self-citations present ([12], [18], [19], [20]) are contextual references to benchmarks and stability studies and are not load-bearing. Honest finding: no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities are postulated. The free parameters are architecture hyperparameters chosen by hand without sensitivity analysis; the axioms reflect standard empirical-validation assumptions and the unverified premise that dynamic gating adds value over static weights.

free parameters (5)
  • Patch size = 8
    Chosen as part of the ViT gating architecture; no sensitivity analysis is reported.
  • Hidden size (Base) = 384
    Chosen for the 25M-parameter model; no sensitivity analysis.
  • Depth (Base) = 6
    Chosen for the 25M-parameter model.
  • Attention heads (Base) = 6
    Chosen for the 25M-parameter model.
  • MLP ratio = 4.0
    Standard Transformer MLP expansion factor, chosen by convention.
axioms (5)
  • domain assumption ERA5 reanalysis is treated as the ground truth for RMSE computation.
    All errors are computed against ERA5 without discussion of reanalysis uncertainty.
  • domain assumption The three expert models (Pangu, Aurora, FCN3) are used as fixed, pre-trained forecasters with complementary errors.
    The approach assumes diversity among experts; no analysis of error correlation is provided.
  • domain assumption The year 2015 is representative of forecast skill beyond the training period (1980-2014).
    Evaluation uses a single year; no multi-year or seasonal breakdown is given.
  • standard math Softmax normalization of weights ensures a convex combination; the bias map captures systematic shared errors.
    Equation (1) with softmax is standard; the bias term is learned and unregularized.
  • domain assumption The ViT gating network can learn meaningful spatial and lead-time dependencies from the stacked expert fields.
    This is the core modeling assumption; the paper does not compare against a linear or static-weight baseline.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of MoWE : A Mixture of Weather Experts." pith.science (2026). https://pith.science/paper/B2MQEUZB

@misc{pith2026250909052,
  author       = {Pith},
  title        = {Pith review of: MoWE : A Mixture of Weather Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2MQEUZB}},
  note         = {Machine review of arXiv:2509.09052}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Data-driven weather models have recently achieved state-of-the-art performance, yet progress has plateaued in recent years. This paper introduces a Mixture of Experts (MoWE) approach as a novel paradigm to overcome these limitations, not by creating a new forecaster, but by optimally combining the outputs of existing models. The MoWE model is trained with significantly lower computational resources than the individual experts. Our model employs a Vision Transformer-based gating network that dynamically learns to weight the contributions of multiple "expert" models at each grid point, conditioned on forecast lead time. This approach creates a synthesized deterministic forecast that is more accurate than any individual component in terms of Root Mean Squared Error (RMSE). Our results demonstrate the effectiveness of this method, achieving up to a 10% lower RMSE than the best-performing AI weather model on a 2-day forecast horizon, significantly outperforming individual experts as well as a simple average across experts. This work presents a computationally efficient and scalable strategy to push the state of the art in data-driven weather prediction by making the most out of leading high-quality forecast models.

Figures

Figures reproduced from arXiv: 2509.09052 by Dallas Foster, Dibyajyoti Chakraborty, Mohammad Amin Nabian, Peter Harrington, Romit Maulik, Sanjay Choudhry.

Figure 1
Figure 1. Figure 1: A diagram illustrating the architecture of the Mixture-of-Weather Experts (MoWE). The left panel shows [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: RMSE Comparison for Weather Forecasting Models. The Root Mean Squared Error (RMSE) is plotted against [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: This image displays 2 m temperature (K) for ERA5, forecasts at 6-hour, 24-hour, and 48-hour from various [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: This image displays wind(m/s) at pressure level 500 for ERA5, forecasts at 6-hour, 24-hour, and 48-hour from various models ( FCN3, Aurora, Pangu, and MoWE) along with the learned weights for FCN3, Aurora, and Pangu in the MoWE. The MoWE model’s forecasts appear reasonable and visually consistent with the others. The weight maps indicate that at the 6-hour forecast, the MoWE places a higher weight on the A… view at source ↗
Figure 5
Figure 5. Figure 5: This image displays geopotential (m2/s2 ) at pressure level 500 for ERA5, forecasts at 6-hour, 24-hour, and 48-hour from various models ( FCN3, Aurora, Pangu, and MoWE) along with the learned weights for FCN3, Aurora, and Pangu in the MoWE. The MoWE model’s forecasts appear reasonable and visually consistent with the others. The weight maps indicate that at the 6-hour forecast, the MoWE places a higher wei… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. AdaWeather: Adaptively Mixing Probabilistic Weather Forecasts with Logarithmic Regret

    cs.LG 2026-06 unverdicted novelty 6.0

    AdaWeather adaptively mixes probabilistic weather forecasts and achieves logarithmic regret relative to the best static mixture of experts in hindsight.

Reference graph

Works this paper leans on

31 extracted references · 11 linked inside Pith · cited by 1 Pith paper

  1. [1]

    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 operators.arXiv preprint arXiv:2202.11214, 2022

    Pith/arXiv arXiv 2022

  2. [2]

    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

    Pith/arXiv arXiv 2022

  3. [3]

    Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023. 9 MoWE : A Mixture of Weather Experts

    2023

  4. [4]

    A foundation model for the earth system.Nature, pages 1–8, 2025

    Cristian Bodnar, Wessel P Bruinsma, Ana Lucic, Megan Stanley, Anna Allen, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan A Weyn, Haiyu Dong, et al. A foundation model for the earth system.Nature, pages 1–8, 2025

    2025

  5. [5]

    Scaling transformers for skillful and reliable medium-range weather forecasting

    Tung Nguyen, Rohan Shah, Hritik Bansal, Troy Arcomano, Sandeep Madireddy, Romit Maulik, Veerabhadra Kotamarthi, Ian Foster, and Aditya Grover. Scaling transformers for skillful and reliable medium-range weather forecasting. InICLR 2024 Workshop on AI4DifferentialEquations In Science, 2024

    2024

  6. [6]

    Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead

    Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv preprint arXiv:2304.02948, 2023

    arXiv 2023

  7. [7]

    Fuxi: a cascade machine learning forecasting system for 15-day global weather forecast.npj climate and atmospheric science, 6(1):190, 2023

    Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: a cascade machine learning forecasting system for 15-day global weather forecast.npj climate and atmospheric science, 6(1):190, 2023

    2023

  8. [8]

    Neural general circulation models for weather and climate.Nature, 632(8027):1060–1066, 2024

    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

    2024

  9. [9]

    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

    Pith/arXiv arXiv 2023

  10. [10]

    Aifs-crps: ensemble forecasting using a model trained with a loss function based on the continuous ranked probability score.arXiv preprint arXiv:2412.15832, 2024

    Simon Lang, Mihai Alexe, Mariana CA Clare, Christopher Roberts, Rilwan Adewoyin, Zied Ben Bouallègue, Matthew Chantry, Jesper Dramsch, Peter D Dueben, Sara Hahner, et al. Aifs-crps: ensemble forecasting using a model trained with a loss function based on the continuous ranked probability score.arXiv preprint arXiv:2412.15832, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Fourcastnet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale.arXiv preprint arXiv:2507.12144, 2025

    Boris Bonev, Thorsten Kurth, Ankur Mahesh, Mauro Bisson, Jean Kossaifi, Karthik Kashinath, Anima Anand- kumar, William D Collins, Michael S Pritchard, and Alexander Keller. Fourcastnet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale.arXiv preprint arXiv:2507.12144, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Analyzing and exploring training recipes for large-scale transformer-based weather prediction.Artificial Intelligence for the Earth Systems, 4(2):240061, 2025

    Jared D Willard, Peter Harrington, Shashank Subramanian, Ankur Mahesh, Travis A O’Brien, and William D Collins. Analyzing and exploring training recipes for large-scale transformer-based weather prediction.Artificial Intelligence for the Earth Systems, 4(2):240061, 2025

    2025

  13. [13]

    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 Modeling Earth Systems, 16(6):e2023MS004019, 2024

    2024

  14. [14]

    On some limitations of current machine learning weather prediction models.Geophysical Research Letters, 51(12):e2023GL107377, 2024

    Massimo Bonavita. On some limitations of current machine learning weather prediction models.Geophysical Research Letters, 51(12):e2023GL107377, 2024

    2024

  15. [15]

    Zied Ben Bouallegue, 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...

    2024

  16. [16]

    Advancing parsimonious deep learning weather prediction using the healpix mesh.Journal of Advances in Modeling Earth Systems, 16(8):e2023MS004021, 2024

    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 prediction using the healpix mesh.Journal of Advances in Modeling Earth Systems, 16(8):e2023MS004021, 2024

    2024

  17. [17]

    Ace: A fast, skillful learned global atmospheric model for climate prediction.arXiv preprint arXiv:2310.02074, 2023

    Oliver Watt-Meyer, Gideon Dresdner, Jeremy McGibbon, Spencer K Clark, Brian Henn, James Duncan, Noah D Brenowitz, Karthik Kashinath, Michael S Pritchard, Boris Bonev, et al. Ace: A fast, skillful learned global atmospheric model for climate prediction.arXiv preprint arXiv:2310.02074, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Towards stability of autoregressive neural operators.arXiv preprint arXiv:2306.10619, 2023

    Michael McCabe, Peter Harrington, Shashank Subramanian, and Jed Brown. Towards stability of autoregressive neural operators.arXiv preprint arXiv:2306.10619, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    A practical probabilistic benchmark for ai weather models.Geophysical Research Letters, 52(7):e2024GL113656, 2025

    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.Geophysical Research Letters, 52(7):e2024GL113656, 2025

    2025

  20. [20]

    Huge ensembles part ii: properties of a huge ensemble of hindcasts generated with spherical fourier neural operators.arXiv preprint arXiv:2408.01581, 2024

    Ankur Mahesh, William Collins, Boris Bonev, Noah Brenowitz, Yair Cohen, Peter Harrington, Karthik Kashinath, Thorsten Kurth, Joshua North, Travis OBrien, et al. Huge ensembles part ii: properties of a huge ensemble of hindcasts generated with spherical fourier neural operators.arXiv preprint arXiv:2408.01581, 2024. 10 MoWE : A Mixture of Weather Experts

    Pith/arXiv arXiv 2024

  21. [21]

    Skillful joint probabilistic weather forecasting from marginals.arXiv preprint arXiv:2506.10772, 2025

    Ferran Alet, Ilan Price, Andrew El-Kadi, Dominic Masters, Stratis Markou, Tom R Andersson, Jacklynn Stott, Remi Lam, Matthew Willson, Alvaro Sanchez-Gonzalez, et al. Skillful joint probabilistic weather forecasting from marginals.arXiv preprint arXiv:2506.10772, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Mark DeMaria, James L Franklin, Galina Chirokova, Jacob Radford, Robert DeMaria, Kate D Musgrave, and Imme Ebert-Uphoff. An operations-based evaluation of tropical cyclone track and intensity forecasts from artificial intelligence weather prediction models.Artificial Intelligence for the Earth Systems, 1(aop), 2025

    2025

  23. [23]

    Predictability limit of the 2021 pacific northwest heatwave from deep- learning sensitivity analysis.Geophysical Research Letters, 51(19):e2024GL110651, 2024

    P Trent V onich and Gregory J Hakim. Predictability limit of the 2021 pacific northwest heatwave from deep- learning sensitivity analysis.Geophysical Research Letters, 51(19):e2024GL110651, 2024

    2021

  24. [24]

    European Centre for Medium Range Weather Forecasts, 2018

    Roberto Buizza, M Alonso Balmaseda, Andrew Brown, S English, Richard Forbes, Alan Geer, T Haiden, Martin Leutbecher, L Magnusson, Mark Rodwell, et al.The development and evaluation process followed at ECMWF to upgrade the Integrated Forecasting System (IFS). European Centre for Medium Range Weather Forecasts, 2018

    2018

  25. [25]

    Young-Youn Park, Roberto Buizza, and Martin Leutbecher. Tigge: Preliminary results on comparing and combining ensembles.Quarterly Journal of the Royal Meteorological Society: A journal of the atmospheric sciences, applied meteorology and physical oceanography, 134(637):2029–2050, 2008

    2029

  26. [26]

    Jonathan A Weyn, Dale R Durran, and Rich Caruana. Can machines learn to predict weather? using deep learning to predict gridded 500-hpa geopotential height from historical weather data.Journal of Advances in Modeling Earth Systems, 11(8):2680–2693, 2019

    2019

  27. [27]

    Evaluation of five global ai models for predicting weather in eastern asia and western pacific.npj Climate and Atmospheric Science, 7(1):221, 2024

    Cheng-Chin Liu, Kathryn Hsu, Melinda S Peng, Der-Song Chen, Pao-Liang Chang, Ling-Feng Hsiao, Chin-Tzu Fong, Jing-Shan Hong, Chia-Ping Cheng, Kuo-Chen Lu, et al. Evaluation of five global ai models for predicting weather in eastern asia and western pacific.npj Climate and Atmospheric Science, 7(1):221, 2024

    2024

  28. [28]

    Piggycast: Improving weather prediction

    Jane Doe. Piggycast: Improving weather prediction. https://thedataandaiteacher.substack.com/p/ piggycast-improving-weather-prediction?triedRedirect=true, August 2025

    2025

  29. [29]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  30. [30]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  31. [31]

    Perceiver io: A general architecture for structured inputs & outputs.arXiv preprint arXiv:2107.14795, 2021

    Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs.arXiv preprint arXiv:2107.14795, 2021. 11

    Pith/arXiv arXiv 2021

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.