Machine-learning correction improves week-three and week-four weather forecasts in tests

Researchers writing in Nature say their lightweight probabilistic bias correction method learns from past forecasts and observations to correct systematic errors in physics-based and AI weather models. They report improved week-three and week-four forecasts, including for extreme weather. Their MicroDuet entry combined corrected forecasts and ranked first for every variable and lead time in the 2025 AI Weather Quest’s September–November season. The researchers say the method could aid farming and disaster planning.