AI‑driven breakthrough doubles accuracy of 2‑to‑6‑week weather forecasts

Researchers from MIT and Microsoft developed Probabilistic Bias Correction (PBC), a machine-learning method that corrects systematic errors in weather models. This technique has successfully doubled the accuracy of 2-to-6-week sub-seasonal weather forecasts when applied to ECMWF models.

The 2-to-6-week forecast window has historically been a weak point in meteorology, hindering planning for agriculture, energy, and disaster management. PBC addresses this by providing a low-cost, CPU-only solution that improves predictive skill without requiring supercomputing resources.

The new method marks the first time an AI-driven approach has consistently outperformed traditional climatology for sub-seasonal forecasts. Additionally, its low-cost design shifts accessibility, potentially narrowing the forecasting gap for resource-constrained regions.