West-WRF AI 2-km: High-Resolution Prediction of Integrated Vapor Transport and Precipitation
Nazak Rouzegari, Vesta Afzali Gorooh, Agniv Sengupta, Phu Nguyen, Kuo-Lin Hsu, Amir AghaKouchak, Soroosh Sorooshian, F. Martin Ralph, et al.
physics.ao-ph · cs.AI
Quick verdict
Claims first 2-km stretched-grid AI forecast skill for western-US AR precipitation/IVT; evidence is broad and honest but confounded and short-period.
Summary
The authors present West-WRF AI 2-km, a deterministic stretched-grid AI forecasting system built on ECMWF's Anemoi Graph Transformer that provides roughly 2-km resolution over the western U.S. and Northeast Pacific within a ~31-km global grid. A global model pretrained on 40 years of ERA5 is fine-tuned on the CW3E 2-km WRF reanalysis (2012–2020) across three training stages, then run autoregressively to 144 hours. Over winters 2020–2023, its precipitation and integrated vapor transport (IVT) forecasts are verified against PRISM 4-km gridded observations, 526 rain gauges, and roughly 2,360 AR Reconnaissance dropsondes, and benchmarked against the West-WRF AI 6-km predecessor, dynamical West-WRF 3-km/9-km, and IFS 9-km. The 2-km model lowers 4-day precipitation RMSE to 0.96 mm/day (15% below IFS 9-km, 4% below dynamical West-WRF 3-km), reproduces fine-scale spectra and extreme-quantile intensities, wins the December 2022 AR3 case study (RMSE 9.64 vs 13.38 mm), and posts the lowest IVT errors at long leads (183.30 kg m⁻¹ s⁻¹ at 120 hours) with better extreme-threshold categorical skill. Its advantage is concentrated in localized, terrain-sensitive extremes; on domain means it merely matches the 6-km predecessor. Limitations: only three winters evaluated, a confounded 2-km-versus-6-km comparison (outer domains differ), reanalysis initialization favoring AI models over operational NWP, and no released model code or weights.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● 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
- West-WRF AI 2-km is, to the authors' knowledge, the first AI forecasting system evaluated specifically for atmospheric-river prediction at ~2-km resolution over the western U.S.
- Refining from 6 km to 2 km preserves large-scale and short-lead skill while adding value concentrated in localized, terrain-sensitive precipitation extremes.
- The 2-km model cuts 4-day PRISM RMSE to 0.96 mm/day (15.0% below IFS 9-km, 4.0% below dynamical West-WRF 3-km) and reduces the Dec 31, 2022 AR3 case RMSE by ~28% versus its 6-km predecessor.
- Against ~2,360 independent dropsondes it achieves the lowest IVT RMSE at every lead time (183.30 kg m⁻¹ s⁻¹ at 120 h) and better categorical skill at the extreme 750 kg m⁻¹ s⁻¹ threshold.
Strengths
Triple, partly model-independent verification
Forecasts are verified against PRISM 4-km (plus 800-m sensitivity, Sec 3a), 526 quality-controlled CNRFC gauges with station-specific 95th-percentile thresholds (Sec 3b), and ~2,360 AR Recon dropsondes computed directly from observed wind/moisture profiles (Sec 2e), which the authors correctly note reduces shared-bias risk relative to reanalysis-based references.
Candid reporting of where the 2-km model does not win
The Discussion explicitly concedes parity with the 6-km predecessor on domain means and short-lead IVT, a 120-hour gauge CSI reversal (0.135 vs 0.186), an AR precipitation-ratio underestimate of ~9.7% (vs −1.1% for 6-km), a +17.7% rainy-day-frequency inflation, and the MSE-loss double-penalty mechanism — unusual transparency for a systems paper.
Above-average bootstrap design
Cleanest evidence is at the operationally critical extreme
Full computational accounting
Weaknesses
Central attribution is confounded
Short, seasonally narrow evaluation
Reanalysis-initialization advantage unquantified
Unexplained regressions on two distributional metrics
Reproducibility is limited
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This preprint (under review) introduces West-WRF AI 2-km, a deterministic stretched-grid extension of CW3E's AI 6-km system (Baño-Medina et al. 2025) built on the ECMWF Anemoi Graph Transformer. Its claims: (1) the first AI forecasting system evaluated specifically for AR prediction at ~2-km resolution over the western U.S. (Introduction); (2) improved localized, terrain-sensitive precipitation extremes and intense AR moisture transport at unchanged large-scale skill; (3) dropsonde-verified IVT advantages that grow with lead time and threshold. The stated objective is a resolution-sensitivity study, not a new architecture — an honest and appropriate framing.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 West-WRF AI 2-km?
It is an AI weather-forecasting model that runs at about 2-km resolution over the western U.S. and Northeast Pacific inside a ~31-km global grid. Built on ECMWF's public Anemoi framework, it produces 6-hourly precipitation and integrated vapor transport (IVT) forecasts out to 6 days (144 hours) autoregressively, in a few minutes on a single H100 GPU.
Does the 2-km AI model beat existing forecasting systems?
It beats the global IFS 9-km model and the dynamical West-WRF on many precipitation metrics (e.g., 4-day RMSE of 0.96 mm/day, 15% lower than IFS), and it is strongest for localized, terrain-sensitive extremes and long-lead IVT against independent dropsondes. Against its own 6-km AI predecessor it is roughly tied on broad-scale metrics and domain means; its clear wins are fine-scale structure, extreme quantiles, a December 2022 AR3 case (RMSE 9.64 vs 13.38 mm), and extreme-threshold IVT detection.
What data was it trained on and how was it evaluated?
It was pretrained on about 40 years of global ERA5 reanalysis, then fine-tuned on eight winters (2012–2020) of the CW3E 2-km WRF-based regional reanalysis. Winters 2020–2023 were held out for evaluation against PRISM gridded precipitation, 526 rain gauges, and roughly 2,360 AR Reconnaissance dropsondes, with benchmarks including West-WRF AI 6-km, dynamical West-WRF 3-km/9-km, and ECMWF IFS 9-km.
Is the model or its code publicly available?
Not fully. The paper points to the public Anemoi framework and public inputs (ERA5, PRISM), but the trained weights and model-specific code are not released, and the CW3E 2-km reanalysis is available only on request from a named CW3E contact — so independent replication is currently not possible.
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 and tables could not be analyzed directly; all numeric results were taken from captions and body text, which quote them densely and consistently.
- Supplemental Material (Figs S1–S9, Tables S1–S2) is not included; hyperparameter sensitivity and the 800-m PRISM results cannot be checked.
- The extracted author list includes Luca Delle Monache, who is absent from the supplied metadata; minor discrepancy noted.
- Equation typography is garbled by extraction (Eqs. 1–8), though the metric definitions remain recoverable.
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.