Researchers improve weather forecasts looking two to six weeks out

Researchers at MIT and collaborators introduced a machine‑learning model that learns to correct errors in existing dynamical forecasts, doubling forecasting accuracy for the two‑to‑six‑week horizon and achieving first place in a major international forecasting competition. The approach also beats climatologic averages for that range and improves prediction of extreme events such as floods and cold spells. For engineers, this demonstrates that post‑processing AI corrections can extend reliable forecast windows well beyond the traditional two‑week limit.

⚡ Key Takeaways

• Forecasting accuracy over the two‑to‑six‑week horizon was doubled when the ML correction model was applied to a gold‑standard dynamical model.
• The ML model operates as an error‑correction layer on top of existing weather forecasting systems.
• It outperforms climatologic averages for the same period, offering more reliable medium‑range predictions.
• Integration involves feeding the output of a standard dynamical model into the trained correction model before final use.
• The method currently relies on a “gold‑standard” forecasting model as input; without such a baseline, the correction model cannot be applied.

Extending reliable forecasts to six weeks enables energy operators, farmers, and emergency responders to plan logistics and mitigate impacts of severe weather far earlier than previously possible.

✅ Practical Steps

1. Obtain forecasts from a high‑quality dynamical weather model and input them into the trained ML correction model as described in the Nature paper.
2. Use the corrected outputs for medium‑range planning tasks such as energy supply scheduling or agricultural decision‑making.

Want the full story? Read the original article.