EXPLAINER

AI Weather and Climate Forecasting Models

conceptfreshness: volatilesources not yet verified

Machine-learning weather models learn from decades of reanalysis data to produce forecasts far faster than traditional numerical models, with accuracy that is competitive on many standard metrics.

Short answer: Models such as GraphCast and Pangu-Weather (and operational ML systems at forecasting centres) forecast global weather in minutes once trained. They complement physics-based numerical weather prediction; strengths and weaknesses (for example extremes, ensembles, physical consistency) are still being studied.

What is it?

Traditional weather forecasts simulate the atmosphere on supercomputers. AI models learn patterns from past weather records and can produce a forecast much faster. Early results are impressive on common measures, though meteorologists still study how well they handle rare, extreme events.

Also known as: GraphCast, Pangu-Weather, FourCastNet, AIFS, machine learning weather forecast, AI weather prediction, neural weather model, ECMWF AI.

Technical definition

Lam et al. (2023, Science) present GraphCast, a graph neural network operating on a multi-mesh over the globe trained on ECMWF ERA5 reanalysis, reporting that it significantly outperformed the most accurate operational deterministic system (ECMWF HRES) on 90% of 1,380 verification targets in the paper's evaluation. Bi et al. (2023, Nature) present Pangu-Weather using 3-D Earth-specific transformers. Others include FourCastNet (Pathak et al., 2022). Forecasting centres have operationalised ML systems: ECMWF's AIFS Single became operational on 25 February 2025 and its ensemble version (AIFS ENS) on 1 July 2025, running alongside the physics-based IFS (status time-sensitive). Open questions include probabilistic/ensemble forecasting, extreme-event fidelity, long-lead and climate-scale applicability, and dependence on physics-based data assimilation for initial conditions.

Entity type: concept · Freshness class: volatile.

When should I use it?

Relevant to forecasters, energy and logistics users, and researchers.

Implementation patterns

  • Autoregressive rollout from analysis state.
  • Ensemble generation with perturbations or generative models.
  • Verification against observations, not only reanalysis.
  • Hybrid pipelines: physics-based assimilation + ML forecast.

Technologies and frameworks

Commonly associated: GraphCast, Pangu-Weather, FourCastNet, ERA5 reanalysis, ECMWF AIFS. Names are examples of what exists, not recommendations, and projects change quickly.

Practical guide

  1. Choose the target variables and lead times.
  2. Compare against operational baselines.
  3. Verify against observations.
  4. Evaluate extremes and ensembles.
  5. Follow official forecasting centre guidance.

Common failure modes

  • Blurring/smoothing of extremes at long lead times.
  • Unphysical drift in long rollouts.
  • Evaluation only on reanalysis rather than observations.
  • Overstating operational readiness.

Troubleshooting

  • Poor extreme-event skill — Evaluate event-specific metrics; consider ensembles or specialised training.
  • Drift at long lead times — Limit rollout length or use constraints and recalibration.

Contrasted with

Questions this page answers

  • How does GraphCast work?
  • Can AI predict the weather better than traditional models?
  • What is AIFS?
  • Do AI weather models work for extreme events?

Typical goals: understand how AI forecasts compare with physics models; assess ML forecasting claims.

Sources and verification

Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.

Freshness class: volatile — Products, versions, protocols, benchmarks, leaderboards, regulations and dates. Verify against the primary source before relying on specifics; re-review every 3 months.

Time-sensitive claims to verify before relying on them:

  • Operational status of ML forecast systems at national centres.
  • Skill comparisons versus numerical models, which change as both improve.

Canonical sources (unverified leads):

  • Lam et al., Learning skillful medium-range global weather forecasting, Science (2023) — arXiv:2212.12794
  • Bi et al., Accurate medium-range global weather forecasting with 3D neural networks, Nature (2023) — doi:10.1038/s41586-023-06185-3
  • Pathak et al., FourCastNet (2022) — arXiv:2202.11214