pith. sign in

arxiv: 2605.28153 · v1 · pith:WZDSPQFYnew · submitted 2026-05-27 · ⚛️ physics.ao-ph · cs.LG

Skillful high-resolution weather forecasting independent of physical models

Pengcheng Zhao , Siqi Xiang , Weixin Jin , Zekun Ni , Jiang Bian , Zuliang Fang , Hongyu Sun , Bin Zhang
show 5 more authors
Richard E. Turner Jonathan Weyn Haiyu Dong Kit Thambiratnam Qi Zhang
This is my paper

Pith reviewed 2026-06-29 09:37 UTC · model grok-4.3

classification ⚛️ physics.ao-ph cs.LG
keywords weather forecastingmachine learningobservational dataregional modelinghigh-resolution forecastsnumerical weather predictionanalysis and prediction
0
0 comments X

The pith

A machine learning system produces skillful high-resolution regional weather forecasts and analyses using only observational data, without any numerical weather prediction models or reanalyses.

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

The paper introduces ObsCast as a regional system that generates both initial conditions and short-term predictions directly from observations. It trains and runs without any NWP-derived data in either stage yet matches or exceeds operational NWP skill for near-surface variables through 18 hours over the contiguous United States and Europe, including for precipitation. This independence removes the need to inherit biases or resolution limits from existing reanalysis products. The approach therefore offers a direct path to build and update regional forecasting services from local observations alone.

Core claim

ObsCast is an end-to-end machine learning model that learns to produce both weather analyses and forecasts solely from observational datasets and achieves state-of-the-art performance in short-term high-resolution regional modeling without using any NWP-generated reanalyses for supervision.

What carries the argument

ObsCast, the regional machine learning system that performs both analysis and forecasting directly from observations.

If this is right

  • Forecasts are generated faster than traditional NWP while maintaining higher skill for near-surface variables.
  • Precipitation forecasts remain skillful without inheriting resolution limits from reanalysis data.
  • Regional services can be built and refined directly from local observations without developing full NWP pipelines.
  • The system adapts more easily to locations where suitable reanalysis products are unavailable or expensive.

Where Pith is reading between the lines

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

  • The same observation-only training approach could be tested on global scales if dense enough observational coverage becomes available.
  • Performance on rare extreme events would indicate whether the model has captured dynamics that physical constraints normally enforce.
  • Combining ObsCast outputs with sparse physical constraints might improve longer-range skill while retaining the independence benefit.

Load-bearing premise

Machine learning models can learn the complex atmospheric dynamics sufficiently well from observational data alone to produce accurate forecasts without guidance from physical models or reanalysis products.

What would settle it

Running ObsCast on a new geographic region with no training observations and comparing its 18-hour forecasts against independent high-resolution measurements would show whether skill falls below that of operational NWP.

Figures

Figures reproduced from arXiv: 2605.28153 by Bin Zhang, Haiyu Dong, Hongyu Sun, Jiang Bian, Jonathan Weyn, Kit Thambiratnam, Pengcheng Zhao, Qi Zhang, Richard E. Turner, Siqi Xiang, Weixin Jin, Zekun Ni, Zuliang Fang.

Figure 0
Figure 0. Figure 0: Impact of training labels on ObsCast-Forecast skill over CONUS. [PITH_FULL_IMAGE:figures/full_fig_p025_0.png] view at source ↗
Figure 0
Figure 0. Figure 0: Robustness of ObsCast-Forecast to reduced station availability at inference [PITH_FULL_IMAGE:figures/full_fig_p026_0.png] view at source ↗
read the original abstract

Accurate and timely weather forecasts are critical for high-impact decisions in modern society. Machine-learning-based weather prediction is emerging as an alternative for producing initial conditions, forecasts, and even both in end-to-end systems. These methods deliver predictions faster and often with higher skill than traditional numerical weather prediction (NWP). However, even end-to-end models typically rely on NWP-generated reanalyses for supervision, thereby inheriting the biases and resolution limitations of those NWPs, and limiting adaptation to settings where suitable reanalysis products are unavailable, infrequently updated, or expensive to produce. Here we introduce ObsCast, a regional system that generates both analysis and predictions, without using any NWP-derived data in either training or inference, while still achieving state-of-the-art performance in short-term high-resolution regional modeling. Over the contiguous United States and Europe, ObsCast outperforms operational NWP for near-surface variables through 18 h and produces skillful precipitation forecasts. It provides a simpler and more adaptable route to build and refine regional forecasting services directly from local observations, without the need to develop complex and costly traditional forecasting pipelines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces ObsCast, a regional machine-learning system that produces both analyses and short-term forecasts at high resolution. It claims to operate entirely from raw observational inputs with no NWP-derived data used in training or inference, while achieving state-of-the-art performance over the contiguous United States and Europe: outperforming operational NWP for near-surface variables through 18 h and delivering skillful precipitation forecasts.

Significance. If the independence claim and quantitative performance results are substantiated with full methodological detail, the work would offer a materially simpler pathway for building regional forecasting systems that do not require access to or inheritance from NWP reanalyses. This could be especially relevant for domains lacking mature NWP infrastructure.

major comments (2)
  1. [Abstract] Abstract: the central performance claims (state-of-the-art skill, outperformance through 18 h, skillful precipitation) are stated without any accompanying metrics, datasets, validation periods, or error bars, rendering the empirical assertions unevaluable from the supplied text.
  2. [Data and Methods] Data and Methods sections: the load-bearing assertion that 'no NWP-derived data' enters training or inference requires an exhaustive accounting of every input field and its provenance. Standard high-resolution observational products (satellite radiances, radar composites, surface QC) frequently embed NWP backgrounds or physical constraints via retrieval or assimilation steps; without explicit verification that ObsCast inputs avoid all such steps, the independence claim cannot be assessed.
minor comments (1)
  1. [Abstract] The abstract and title would benefit from a concise operational definition of 'NWP-derived data' to bound the independence claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback and recommendation for major revision. We address each major comment below and will revise the manuscript accordingly to improve clarity and substantiation of our claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claims (state-of-the-art skill, outperformance through 18 h, skillful precipitation) are stated without any accompanying metrics, datasets, validation periods, or error bars, rendering the empirical assertions unevaluable from the supplied text.

    Authors: We agree that the abstract would benefit from greater specificity. In the revised manuscript we will incorporate key quantitative metrics (e.g., RMSE and anomaly correlation for 2 m temperature and 10 m wind), the exact validation periods and domains (e.g., 2022–2023 over CONUS and Europe), and reference to error bars or significance testing so that the performance claims can be directly evaluated from the abstract. revision: yes

  2. Referee: [Data and Methods] Data and Methods sections: the load-bearing assertion that 'no NWP-derived data' enters training or inference requires an exhaustive accounting of every input field and its provenance. Standard high-resolution observational products (satellite radiances, radar composites, surface QC) frequently embed NWP backgrounds or physical constraints via retrieval or assimilation steps; without explicit verification that ObsCast inputs avoid all such steps, the independence claim cannot be assessed.

    Authors: We acknowledge that an exhaustive provenance table is necessary to fully substantiate the independence claim. The revised Data and Methods section will include a comprehensive table enumerating every input field, its exact observational source, and explicit confirmation that no NWP background, retrieval, or assimilation step is involved (e.g., direct ASOS surface observations, raw radar reflectivity, and unprocessed satellite brightness temperatures). This will allow readers to verify that no NWP-derived information enters training or inference. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical performance claims from observational data

full rationale

The paper presents ObsCast as an ML system trained and run exclusively on observational inputs to produce analysis and forecasts, with performance evaluated empirically against operational NWP. No equations, derivations, or claims reduce by construction to fitted parameters, self-definitions, or self-citation chains. The independence assertion and skill results are framed as outcomes of data-driven training rather than tautological or load-bearing self-references, making the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract supplies no information on free parameters, axioms, or invented entities; none can be identified.

pith-pipeline@v0.9.1-grok · 5761 in / 1248 out tokens · 37338 ms · 2026-06-29T09:37:04.808712+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

49 extracted references · 9 canonical work pages · 1 internal anchor

  1. [1]

    & Brunet, G

    Bauer, P., Thorpe, A. & Brunet, G. The quiet revolution of numerical weather prediction.Nature525, 47–55 (2015)

    2015

  2. [2]

    Kalnay, E.Atmospheric modeling, data assimilation and predictability(Cam- bridge university press, 2003)

    2003

  3. [3]

    J.et al.All-sky satellite data assimilation at operational weather forecasting centres.Quarterly Journal of the Royal Meteorological Society144, 1191–1217 (2018)

    Geer, A. J.et al.All-sky satellite data assimilation at operational weather forecasting centres.Quarterly Journal of the Royal Meteorological Society144, 1191–1217 (2018)

    2018

  4. [4]

    Reichstein, M.et al.Deep learning and process understanding for data-driven earth system science.Nature566, 195–204 (2019)

    2019

  5. [5]

    Ben Bouallegue, Z.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 American Meteorological Society105, E864–E883 (2024)

    2024

  6. [6]

    Bi, K.et al.Accurate medium-range global weather forecasting with 3d neural networks.Nature619, 533–538 (2023)

    2023

  7. [7]

    Chen, L.et al.Fuxi: a cascade machine learning forecasting system for 15-day global weather forecast.npj climate and atmospheric science6, 190 (2023)

    2023

  8. [8]

    Lam, R.et al.Learning skillful medium-range global weather forecasting.Science 382, 1416–1421 (2023)

    2023

  9. [9]

    Nature632, 1060–1066 (2024)

    Kochkov, D.et al.Neural general circulation models for weather and climate. Nature632, 1060–1066 (2024)

    2024

  10. [10]

    Lang, S.et al.Aifs–ecmwf’s data-driven forecasting system.arXiv preprint arXiv:2406.01465(2024)

    work page arXiv 2024

  11. [11]

    Price, I.et al.Probabilistic weather forecasting with machine learning.Nature 637, 84–90 (2025). 19

    2025

  12. [12]

    Bodnar, C.et al.A foundation model for the earth system.Nature1–8 (2025)

    2025

  13. [13]

    Parker, W. S. Reanalyses and observations: What’s the difference?Bulletin of the American Meteorological Society97, 1565–1572 (2016)

    2016

  14. [14]

    Hersbach, H.et al.The era5 global reanalysis.Quarterly journal of the royal meteorological society146, 1999–2049 (2020)

    1999

  15. [15]

    Chen, K.et al.Towards an end-to-end artificial intelligence driven global weather forecasting system.arXiv preprint arXiv:2312.12462(2023)

    work page arXiv 2023

  16. [16]

    Alexe, M.et al.Graphdop: Towards skilful data-driven medium-range weather forecasts learnt and initialised directly from observations.arXiv preprint arXiv:2412.15687(2024)

    work page arXiv 2024

  17. [17]

    Allen, A.et al.End-to-end data-driven weather prediction.Nature641, 1172– 1179 (2025)

    2025

  18. [18]

    Nature Communications16, 6658 (2025)

    Sun, X.et al.A data-to-forecast machine learning system for global weather. Nature Communications16, 6658 (2025)

    2025

  19. [19]

    Litjens, G.et al.A survey on deep learning in medical image analysis.Medical image analysis42, 60–88 (2017)

    2017

  20. [20]

    Choudhary, K.et al.Recent advances and applications of deep learning methods in materials science.npj Computational Materials8, 59 (2022)

    2022

  21. [21]

    G.et al.Can deep learning beat numerical weather prediction? Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences379(2021)

    Schultz, M. G.et al.Can deep learning beat numerical weather prediction? Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences379(2021)

    2021

  22. [22]

    McNally, A.et al.Data driven weather forecasts trained and initialised directly from observations.arXiv preprint arXiv:2407.15586(2024)

    work page arXiv 2024

  23. [23]

    Manshausen, P.et al.Generative data assimilation of sparse weather station observations at kilometer scales.Journal of Advances in Modeling Earth Systems 17, e2024MS004505 (2025)

    2025

  24. [24]

    Xiang, Y.et al.Adaf: An artificial intelligence data assimilation framework for weather forecasting.Journal of Advances in Modeling Earth Systems17, e2024MS004839 (2025)

    2025

  25. [25]

    Fan, H.et al.Physically consistent global atmospheric data assimilation with machine learning in latent space.Science Advances12, eaea4248 (2026)

    2026

  26. [26]

    T., Peterson, R

    Warner, T. T., Peterson, R. A. & Treadon, R. E. A tutorial on lateral boundary conditions as a basic and potentially serious limitation to regional numerical 20 weather prediction.Bulletin of the American Meteorological Society78, 2599– 2618 (1997)

    1997

  27. [27]

    S.et al.The real-time mesoscale analysis at noaa’s national centers for environmental prediction: current status and development.Weather and Forecasting26, 593–612 (2011)

    De Pondeca, M. S.et al.The real-time mesoscale analysis at noaa’s national centers for environmental prediction: current status and development.Weather and Forecasting26, 593–612 (2011)

    2011

  28. [28]

    Zhang, J.et al.Multi-radar multi-sensor (mrms) quantitative precipitation esti- mation: Initial operating capabilities.Bulletin of the American Meteorological Society97, 621–638 (2016)

    2016

  29. [29]

    C.et al.The high-resolution rapid refresh (hrrr): An hourly updating convection-allowing forecast model

    Dowell, D. C.et al.The high-resolution rapid refresh (hrrr): An hourly updating convection-allowing forecast model. part i: Motivation and system description. Weather and Forecasting37, 1371–1395 (2022)

    2022

  30. [30]

    URL https://www

    ECMWF.IFS Documentation CY49R1(ECMWF, 2024). URL https://www. ecmwf.int/en/elibrary/81623-ifs-documentation-cy49r1-part-i-observations

    2024

  31. [31]

    Noaa global fore- cast system (gfs)

    National Oceanic and Atmospheric Administration (NOAA). Noaa global fore- cast system (gfs). Registry of Open Data on AWS (2024). URL https://registry. opendata.aws/noaa-gfs-bdp-pds/. Accessed: 2026-02-05

    2024

  32. [32]

    Fovell, R. G. & Gallagher, A. An evaluation of surface wind and gust forecasts from the high-resolution rapid refresh model.Weather and Forecasting37, 1049– 1068 (2022)

    2022

  33. [33]

    P.et al.The high-resolution rapid refresh (hrrr): an hourly updating convection-allowing forecast model

    James, E. P.et al.The high-resolution rapid refresh (hrrr): an hourly updating convection-allowing forecast model. part ii: Forecast performance.Weather and Forecasting37, 1397–1417 (2022)

    2022

  34. [34]

    Ravuri, S.et al.Skilful precipitation nowcasting using deep generative models of radar.Nature597, 672–677 (2021)

    2021

  35. [35]

    Nature619, 526–532 (2023)

    Zhang, Y.et al.Skilful nowcasting of extreme precipitation with nowcastnet. Nature619, 526–532 (2023)

    2023

  36. [36]

    Janji´ c, T.et al.On the representation error in data assimilation.Quarterly Journal of the Royal Meteorological Society144, 1257–1278 (2018)

    2018

  37. [37]

    Xu, P.et al.An artificial intelligence-based limited area model for forecasting of surface meteorological variables.Communications Earth & Environment6, 372 (2025)

    2025

  38. [38]

    Dueben, P. D. & Bauer, P. Challenges and design choices for global weather and climate models based on machine learning.Geoscientific Model Development11, 3999–4009 (2018). 21

    2018

  39. [39]

    Jin, W.et al.Weatherreal: a benchmark based on in-situ observations for evaluating weather models.arXiv preprint arXiv:2409.09371(2024)

    work page arXiv 2024

  40. [40]

    NOAA GOES-16 Advanced Baseline Imager (ABI) Data (2019)

    GOES-R Series Program. NOAA GOES-16 Advanced Baseline Imager (ABI) Data (2019). Available at https://www.ncei.noaa.gov/

    2019

  41. [41]

    Meteosat Satellite Observations

    European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT). Meteosat Satellite Observations. EUMETSAT data and prod- ucts accessed via EUMETSAT Data Centre (2025). URL https://www.eumetsat. int/our-satellites/meteosat-series

    2025

  42. [42]

    Saltikoff, E.et al.Opera the radar project.Atmosphere10, 320 (2019)

    2019

  43. [43]

    Weng, F.et al.Advanced microwave sounding unit cloud and precipitation algorithms.Radio Science38, 33–1 (2003)

    2003

  44. [44]

    Habib, S., Meynart, R., Neeck, S

    Bonsignori, R. Habib, S., Meynart, R., Neeck, S. P. & Shimoda, H. (eds)The Microwave Humidity Sounder (MHS): in-orbit performance assessment. (eds Habib, S., Meynart, R., Neeck, S. P. & Shimoda, H.)Sensors, Systems, and Next- Generation Satellites XI, Vol. 6744, 67440A. International Society for Optics and Photonics (SPIE, 2007). URL https://doi.org/10.11...

    work page doi:10.1117/12.737986 2007

  45. [45]

    Durre, I.et al.Integrated global radiosonde archive (igra), version 2 (2016)

    2016

  46. [46]

    Danielson, J. J. & Gesch, D. B. Global multi-resolution terrain elevation data 2010 (gmted2010). Tech. Rep., US Geological Survey (2011)

    2010

  47. [47]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

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

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  48. [48]

    URL https://arxiv.org/abs/2111.09883

    Liu, Z.et al.Swin transformer v2: Scaling up capacity and resolution (2022). URL https://arxiv.org/abs/2111.09883. 2111.09883

    work page arXiv 2022

  49. [49]

    Acknowledgments The authors express their gratitude to all providers for providing the publicly acces- sible data

    Guibas, J.et al.Adaptive fourier neural operators: Efficient token mixers for transformers.arXiv preprint arXiv:2111.13587(2021). Acknowledgments The authors express their gratitude to all providers for providing the publicly acces- sible data. The authors would like to thank Synoptic Data PBC (accessible at https://synopticdata.com/) for aggregating stat...

    work page arXiv 2021