Historic El Niño to test AI weather models' extreme limits
TL;DR
- The Financial Times says AI weather models face a major test under a historic El Niño, with some scientists questioning the technology's accuracy.
- The core worry is the "extrapolation problem": AI trained on 1979–2017 data tends to underestimate record-breaking heat, cold and wind events.
- Peer-reviewed work found physics-based models still outperform AI on record-breaking extremes, even as AI wins on speed and general accuracy.
AI's ability to forecast extreme weather faces a major test in the coming months amid a historic El Niño, the Financial Times reports, with some scientists questioning the technology's accuracy. The paper notes the systems are already being used to anticipate everything from forest fires to futures contracts, with Google DeepMind's GraphCast among the models in widest use.
The concern scientists keep returning to is what they call the "extrapolation problem." AI models learn from recent decades of atmospheric data, and a strong El Niño pushes the atmosphere outside that record.
The FT writes that, compared with a leading physics-based model, "AI systems tend to underestimate the intensity and frequency of record-breaking heat, cold and wind events." That tracks a peer-reviewed study covered by Carbon Brief in which AI systems trained on data from 1979 to 2017 showed difficulty in generalizing beyond the data on which they were trained when tested against record-breaking events.
Our climate coverage tracker has logged 42 AI-and-climate stories in the last 90 days, most of them on the power-and-grid side of the build-out; forecasting itself has not had a stress test like this one.
Originally reported by ft.com
Read the original article →Original headline: FT: AI Weather Models Face Historic El Niño Test as Some Scientists Question Accuracy