AI weather forecasting

The paradigm shift

For 70 years, weather forecasting has meant solving the equations of atmospheric physics on supercomputers. Slow, expensive, effective.

In 2023, Google DeepMind's GraphCast beat the industry standard ECMWF model on many metrics. In 2023, Huawei's Pangu did too. In 2024, Microsoft's Aurora. All run on GPUs in seconds instead of hours on CPUs.

This isn't incremental improvement. It's a new paradigm.

How AI weather models work

1. Start with 40+ years of past weather data (ERA5 reanalysis dataset).
2. Train a neural network to predict future weather from past weather.
3. Not physics — pattern learning.
4. Model learns implicit physics from data.
5. Result: given current atmospheric state, predict state 6 hours later.
6. Chain 6-hour steps to get 10-day forecasts.
7. GPU inference: seconds per forecast.

The three main systems

What they're good at

• Medium-range forecasts (3-10 days).
• Storm track prediction.
• Temperature and pressure forecasting.
• Global-scale patterns.
• Running in seconds instead of hours.
• Being trained on cheaper hardware.
• Ensembling — running many variations quickly.

What they're still bad at

• Extreme events they haven't seen before.
• Fine-scale convection (like individual thunderstorms).
• Precipitation exact amounts and timing.
• Tornado prediction (not enough training data on tornado details).
• Climate-change-driven trends outside training distribution.
• Small-scale wind features.
• Fog.
• Ocean-atmosphere coupling for hurricane intensity.

Why they matter for chasers and safety

• AI models will supplement, not replace, HRRR for chase-day decisions.
• Better medium-range means better multi-day chase planning.
• AI ensemble outputs will show forecast confidence.
• AI-augmented radar for tornado detection is in development.
• Warn-on-Forecast with AI acceleration coming.
• Personal AI weather assistants for individual users.

What NWS does with them

• As of 2026, NWS uses AI models as inputs to human forecast decisions.
• Not as final forecast product.
• Ensemble includes GFS + AI outputs.
• National Blend of Models (NBM) increasingly incorporates AI.
• Human forecaster still makes the call for warnings and outlooks.

The training data problem

• AI models trained on 1979-present data.
• Extreme events (EF5 tornadoes, Cat 5 hurricanes) are RARE in training.
• Result: AI models underforecast extreme events.
• Climate change is producing conditions not in training.
• Training data quality varies globally — poor for developing world.
• Bias toward "average" outcomes.

The interpretability problem

Traditional models: we know why they output what they do (equations).

AI models: we don't. It's a black box.

• Meteorologists can't always trust an AI forecast for a specific event.
• When AI and physics disagree, which is right?
• Verification against ground truth is important.
• AI models can hallucinate — produce physically impossible states.
• This limits how much you can rely on them in operational settings.

The future

• Hybrid models: physics + ML. Best of both.
• Fine-tuned regional models.
• Convection-allowing AI models.
• Continuous learning: models updated with new events.
• Ensemble AI for uncertainty quantification.
• AI for radar interpretation.
• AI for satellite interpretation.
• AI-generated forecast text for public consumption.
• AI for warning coordination.

Learn more

• How forecast accuracy improved
• Reading SPC mesoanalysis
• NOAA data sources
• Warning lead time history