WeatherNext 3: Increasing resolution and performance of global weather models with raw observations
Stephan Rasp, Boris Babenko, Dominic Masters, Andrew El-Kadi, Samier Merchant, Guy Shalev, Ilan Price, Fred Zyda, et al.
cs.LG
Quick verdict
WN3 claims SOTA probabilistic forecasting with hourly satellite inputs, 0.1° output, observation training; strong evaluation but key data proprietary.
Summary
WeatherNext 3 (WN3) is Google DeepMind's operational global AI ensemble forecasting system, built to fix two shortcomings of prior AI weather models: reliance on 6-hourly analyses and limited resolution. WN3 scales the FGN architecture of WN2 (latent 1024, 32 mesh layers) and adds three capabilities: hourly forecast initialization by ingesting low-latency geostationary satellite mosaics; native 0.1°/hourly output of single-level variables; and direct training against observation-derived targets — proprietary PARDIG and IMERG precipitation products, tropical cyclone records, and a continuous station head predicting 2m temperature/dewpoint at any location. In 2024 evaluations, WN3 beats WN2 on upper-air CRPS by ~5% (≈6h extra skill); station-head 2m temperature CRPS improves up to 30% vs WN2 and 40% vs ENS at short leads; precipitation CRPS drops up to 60% (IMERG), 30% (MRMS), 10% (rain gauges). A six-week 2026 real-time comparison shows ~10% upper-air advantage over AIFS ENS v2. The paper candidly reports hexagonal mesh artifacts, 6-hour boundary discontinuities in station output, ensemble under-dispersion, and an unexplained 6h-lead degradation. Limitations: analysis metrics use HRES-fc0, which is also the model's own input (circularity); PARDIG is undisclosed; extreme precipitation is not evaluated; no significance testing; no code release.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- WN3 establishes a new state-of-the-art for probabilistic medium-range forecasting, outperforming WN2 (~5% upper-air CRPS, ≈6h extra skill) and AIFS ENS v2 (~10% in a 6-week 2026 real-time evaluation).
- Ingesting low-latency geostationary satellite data enables hourly forecast initialization, delivering a 2-3 hour operational lead-time gain that most benefits fast-evolving precipitation.
- Training directly against observation-derived targets (satellite precipitation, cyclones, stations) yields substantially lower error than analysis-trained models, with station-head 2m temperature CRPS up to 30% better than WN2 and 40% better than ENS even at held-out stations.
- The station head produces geographically-informed 2m temperature/dewpoint at arbitrary locations and times, eliminating the separate statistical post-processing stage used in traditional pipelines.
Strengths
Training beyond analyses is a genuine conceptual advance
Section 2.3/Table 1: direct targets include satellite precipitation (PARDIG, IMERG), IBTrACS cyclones, and sparse station data; the continuous station head (Eq. A.3) interpolates a 0.1° latent with elevation/land-sea/time-offset metadata and generalizes to unseen locations — overfitting to training stations is only a few percent (Fig. A.5), validated on a 5% held-out METAR/Mesonet split.
Latency-adjusted evaluation isolates the value of hourly initialization
Section 3.3/Fig. 4c/A.7: comparing identical models on hourly vs 6-hourly init cycles under an assumed 7h operational latency yields a measured 2-3 hour operational lead gain, matching the theoretical ~3h average — a clean ablation-style design rarely seen in this literature.
Unusually candid artifact and failure reporting
Multi-ground-truth verification of precipitation claims
Full operational realism and cost transparency
Weaknesses
Circular ground truth for headline analysis scores
Headline precipitation gains depend on an undisclosed proprietary product
No statistical uncertainty on any comparison
Unexplained failure modes left without ablations
Extreme precipitation and reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
WeatherNext 3 (WN3), from Google DeepMind and Google Research, is an operational 15-day, 64-member global ensemble AI forecasting system presented as a new state-of-the-art in probabilistic medium-range prediction. Three headline claims structure the paper: (1) hourly forecast initialization by ingesting low-latency geostationary satellite mosaics (Section 2.2); (2) native 0.1°/hourly output of single-level variables, adding solar radiation and cloud cover; and (3) moving beyond analysis-only training via satellite precipitation products (proprietary PARDIG plus IMERG), tropical cyclone attributes, and a continuous station head for 2m temperature/dewpoint queryable anywhere on Earth. The paper is an empirical system paper continuing the GraphCast → GenCast → WN2 lineage (Price et al. 2025; Alet et al. 2025).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is WeatherNext 3?
WeatherNext 3 (WN3) is an operational AI weather forecasting system from Google DeepMind. It issues 15-day, 64-member global ensemble forecasts every hour (rather than every 6 hours), reaches 0.1° resolution for surface variables, and is trained not only on analysis data but directly on observations: satellite-derived precipitation, tropical cyclone records, and weather-station measurements.
How much better is WeatherNext 3 than previous models?
In full-year 2024 evaluations, WN3 beats its predecessor WN2 by about 5% in upper-air CRPS (equivalent to roughly 6 hours of extra forecast skill), improves 2m temperature error at unseen weather stations by up to 30% versus WN2 and 40% versus the ECMWF ensemble at short lead times, and cuts precipitation CRPS by up to 60% (vs IMERG), 30% (vs MRMS) and 10% (vs rain gauges). In a six-week 2026 real-time test it led ECMWF's AIFS ENS v2 by about 10% on upper-air variables in the first week.
Can I access WeatherNext 3 forecasts or its code?
Forecasts are available through deepmind.google/science/weathernext/. The model code, weights, and training data are not released; notably, the PARDIG precipitation dataset central to the headline results is proprietary and its methodology had not been published as of this paper.
What are the known weaknesses of WeatherNext 3?
The authors themselves report: hexagonal mesh-shaped artifacts in individual precipitation and station forecasts, temporal jumps at 6-hour boundaries in station output, per-member global temperature biases, under-dispersive tropical cyclone ensembles, an unexplained skill dip at 6-hour lead times, and no evaluation of extreme precipitation above 4 mm per 6 hours. Ensemble statistics like the median remain usable.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures appear in the extracted text only as captions; plotted scorecards, reliability diagrams and artifact panels could not be inspected directly, so figure-level judgments rest on captions and in-text numbers.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.