AI Tools for Climate & Earth Science
AI weather models have started to match or beat physics-based forecasts at a fraction of the compute cost — a guide to the operational and research tools reshaping the field.
Why this field matters right now
Weather and climate modeling has traditionally relied on physics-based numerical weather prediction (NWP) — computationally expensive simulations run on supercomputers. AI weather models trained on decades of reanalysis data have started to match or beat those physics-based models on many metrics, while running dramatically faster and cheaper. This is one of the few AI-for-science areas where models are already operational at national meteorological agencies, not just research prototypes — making it a good field to point to when researchers ask whether AI-for-science claims are real or overhyped.
Core tools
GraphCast
What it does: A graph neural network model from Google DeepMind that forecasts global weather by representing the Earth as a mesh of connected nodes and passing information between them across multiple spatial scales. It generates 10-day global forecasts and outperformed ECMWF’s best deterministic model on 90% of 1,380 test metrics in its original benchmark.
Access: Open research release; also available as a hindcast archive (UT Austin’s GraphCast Hindcast dataset covers 1979–2024) for researchers who want a ready-made historical dataset rather than running the model themselves.
Best for: Medium-range (1–10 day) global weather forecasting research, and as a fast baseline to compare new methods against.
Note: GraphCast is a research model, not a hosted product — expect to work with code/data releases rather than a sign-up flow.
Pangu-Weather
What it does: A transformer-based global weather forecasting model from Huawei, offering an alternative architecture to GraphCast’s graph-neural-network approach. In its original benchmarks it ran 10,000× faster than traditional ensemble numerical weather prediction.
Access: Open research release.
Best for: Comparing against GraphCast if you want to understand how architecture choice (transformer vs. graph neural network) affects forecast behavior.
GenCast
What it does: A follow-up DeepMind model using a diffusion-based approach to generate ensemble (probabilistic) forecasts rather than a single deterministic prediction — useful for quantifying forecast uncertainty, which GraphCast alone does not provide. Published in Nature (December 2024), GenCast beat ECMWF’s operational ensemble system (ENS) on 97.2% of 1,320 verification targets, with the win rate climbing to 99.8% at lead times beyond 36 hours.
Access: Open — DeepMind released code, weights, and forecasts publicly.
Best for: Work requiring uncertainty quantification, such as risk assessment or decision-support applications where a single forecast isn’t enough.
Note: A newer open-source competitor, NVIDIA’s Earth-2 suite, claims to outperform GenCast, but that comparison isn’t yet independently peer-reviewed — worth tracking rather than treating as settled. See our news coverage for details.
ECMWF AIFS
What it does: The European Centre for Medium-Range Weather Forecasts’ own operational AI model — notable because ECMWF moved AIFS to operational status in 2024, making it the first major meteorological agency to run an AI model operationally rather than as a research add-on.
Access: Operational at ECMWF; research access varies by data-sharing agreement.
Best for: Researchers wanting to study or benchmark against an agency-operational model rather than a purely academic release.
NOAA AIGFS / AIGEFS
What it does: NOAA’s own AI forecast systems, built in part on a GraphCast foundation fine-tuned with NOAA’s own data assimilation system. This additional fine-tuning improved on the base model’s performance, particularly when using NOAA’s own initial conditions, and the ensemble version has shown improved performance over NOAA’s traditional ensemble system, extending forecast skill by an additional 18 to 24 hours.
Access: Operational within NOAA; research access and documentation available through NOAA’s Earth Prediction Innovation Center.
Best for: U.S.-focused forecasting research, or as an example of how agencies fine-tune general models like GraphCast on their own data.
Suggested workflow
| Step | Tool | Purpose |
|---|---|---|
| 1 | GraphCast (or the UT Austin hindcast archive) | Establish a fast, strong baseline forecast |
| 2 | GenCast | Add ensemble/uncertainty quantification if your application needs it |
| 3 | ECMWF AIFS / NOAA AIGFS output | Compare against an agency-operational model for real-world validation |
| 4 | Domain-specific physics validation | Confirm results hold up in your specific use case (e.g., hurricane tracking, precipitation extremes) |
What these tools can’t do yet
- Extreme and short-range events remain a weak spot. AI weather models’ primary weaknesses as of 2026 remain extreme precipitation, short-range convection, and performance in climate-shifted conditions outside their training data — don’t assume a model that excels at medium-range forecasts will be equally reliable for a sudden convective storm.
- Pure AI isn’t replacing physics-based modeling — it’s merging with it. The field’s own practitioners expect the future to belong to hybrid models that combine machine learning’s pattern recognition with physics-based constraints, not pure AI replacing numerical weather prediction entirely. Frame your own work with that trajectory in mind rather than treating AI and physics-based NWP as competitors.
- Training data cutoffs matter for climate work specifically. Since these models are trained on historical reanalysis data (commonly ERA5, 1979 onward), using them for climate-shifted future scenarios outside that historical distribution requires real caution — this is a different failure mode than weather forecasting on a normal day.
Related content
- Glossary: “Reanalysis data,” “Numerical weather prediction (NWP),” “Ensemble forecasting,” “Graph neural network”
- Field Guide: Materials Science — for a comparison of how AI-driven prediction plays out in another physics-heavy field
- Field Guide: Structural Biology