AI Weather Models Are Coming for Traditional Forecasting

The Death of the Supercomputer Monopoly

For sixty years, weather forecasting has been dominated by numerical weather prediction (NWP) — massive systems of partial differential equations solved on supercomputers costing hundreds of millions of dollars. ECMWF's IFS, NOAA's GFS, and MeteoFrance's Arpège have been the gold standard.

In 2023, that paradigm started cracking.

The AI Challengers

GraphCast (Google DeepMind)

GraphCast uses a graph neural network (GNN) trained on 39 years of ERA5 reanalysis data. It produces 10-day forecasts in under 60 seconds on a single TPU — a task that takes ECMWF's IFS hours on a supercomputer cluster.

The results? GraphCast outperformed HRES (ECMWF's high-resolution model) on 90% of 1,380 verification targets.

Pangu-Weather (Huawei)

Pangu-Weather uses a 3D vision transformer architecture. It treats the atmosphere as a 3D image and applies self-attention across pressure levels, latitude, and longitude. It achieves comparable accuracy to IFS with inference times measured in seconds.

FourCastNet (NVIDIA)

NVIDIA's approach uses Adaptive Fourier Neural Operators (AFNOs) — essentially learning the physics in spectral space rather than grid space. It's particularly good at capturing extreme events.

Why This Matters for Geoscience

The implications are profound:

1. Democratization: A $10K GPU can now produce forecasts that previously required $100M infrastructure
2. Ensemble scaling: You can run 1000 ensemble members in the time it took to run 50
3. Tropical cyclone tracking: AI models are showing particular strength in TC track prediction
4. Sub-seasonal prediction: The 2-4 week gap is where AI might make the biggest gains

The Catch

AI weather models are interpolators, not simulators. They learn statistical patterns from historical data. They cannot:

• Predict truly unprecedented events (climate change extremes)
• Provide physical explanations for their forecasts
• Handle novel climate regimes they weren't trained on

The future isn't AI replacing NWP — it's AI and NWP working together. Hybrid approaches are already showing the most promise.

Getting Started

If you're a geoscientist interested in AI weather models, start with:

# Install WeatherBench2 for evaluation
pip install weatherbench2

# Download ERA5 data via CDS API
pip install cdsapi

The field is moving fast. Papers are dropping weekly. The best time to enter AI weather prediction was 2023. The second best time is now.

---

I'm pursuing research at the intersection of AI and geoscience. Follow this blog for more deep dives.