AI for Climate: How Foundation Models Are Improving Extreme Weather Prediction

AI for Climate: How Foundation Models Are Improving Extreme Weather Prediction

GraphCast and Pangu-Weather delivered measurable gains in 10-day forecast accuracy during 2025 tests. Traditional numerical models still lead on some extremes, forcing agencies like the National Weather Service to blend approaches. The shift matters for planners who need reliable signals on floods, heat waves, and storms.

Foundation models treat the atmosphere as a learned pattern rather than a set of physics equations solved step by step. This difference produces faster runs and sometimes sharper detail at longer ranges. Accuracy gains are clearest on mid-latitude temperature and wind, while heavy precipitation remains harder to pin down.

[Model Architectures]

• GraphCast: Graph neural network trained on decades of ERA5 reanalysis data

• Pangu-Weather: Transformer-based system that separates vertical atmospheric layers for efficiency

• Traditional models: Solve partial differential equations on global grids at set time steps

NWS began limited operational testing of AI outputs in early 2026 alongside its existing Global Forecast System. Forecasters review both streams and issue blended products for certain variables. Full replacement is not planned yet.

Heavy precipitation still shows larger errors in AI runs, especially when events depend on small-scale convection that reanalysis data does not fully capture. Operational centers therefore keep physics-based ensembles active for these cases.

Adaptation plans now treat improved 7-to-10-day skill as a planning input rather than a curiosity. Utilities and emergency offices can stage resources earlier when AI signals align with numerical guidance. Disagreement between the two systems still triggers extra review.

Future work focuses on training with higher-resolution satellite and radar observations to close the precipitation gap. Hybrid systems that correct AI output with physical constraints are under active development at several centers.

The next clear milestones will be the 2026 hurricane season results and any published verification scores for summer convective events. Planners watch those periods for evidence that the accuracy edge holds on high-impact extremes.