Researchers testing Pangu-Weather, an AI model for global weather forecasts, found that changing its starting weather data could send Typhoon Khanun along two sharply different paths: one turned east near the Ryukyu Islands, while another continued northwest toward China. In a test that changed data in two regions at once, the model’s predicted position on day seven was 1,006 kilometers from the forecast that used unchanged starting data.
Weather forecasters can run an ensemble: many versions of the same forecast, each starting from slightly different atmospheric conditions. The spread shows how sensitive a prediction is to what the model knows at the start. For Khanun, the researchers used ERA5, a historical global weather dataset, to create 528 starting states with changes constrained by atmospheric physics.
The resulting forecasts formed three groups. Fifty-eight percent turned east past the Ryukyu Islands, 24 percent continued northwest toward China, and 18 percent went north or followed other tracks. Only the combined changes in two regions produced the China-bound shift; changing either region alone did not.
The authors traced the shift to winds that steer a cyclone. In the combined test, a stronger high-pressure system north of Khanun and weaker southwesterly monsoon winds changed the steering flow enough to suppress the predicted eastward turn near the Ryukyu Islands. This is a model sensitivity result, not evidence that a live forecast service made a 1,006-kilometer error.
Separately, the team found rapid track revisions in 14 of 112 western North Pacific cyclones in Pangu-Weather forecasts of past cyclones from 2020 through 2024. That broader count is distinct from the 528-run Khanun experiment; the paper does not report the same perturbation test for every cyclone.
The study is a September 30 arXiv preprint and tests one AI model with historical data. It shows that some physically constrained changes to the starting conditions can split an AI forecast into different tracks, but it does not establish that other models behave the same way. In the researchers’ smallest one-hour perturbation test, only one ensemble member took the alternate northwestward path; the larger splits appeared at stronger settings.
Sources
- Butterfly Effect Confirmed in Global AI Weather Models: Evidence from Tropical Cyclone Forecasting (arXiv preprint, September 30, 2026)
- A spectral test of the butterfly effect and physical consistency in the diffusion-based GenCast’s ensembles (npj Climate and Atmospheric Science, March 18, 2026)
- Image: NASA Earth Observatory, using NOAA-20 VIIRS data; image and event details on NASA’s source page.