AI improves weather forecasting: Researchers double reliability for forecasts 2 to 6 weeks ahead

Current weather forecasts are relatively reliable for periods of up to 10 days. Temperatures, rainfall, wind, and the timing of individual weather patterns can often be predicted with reasonable accuracy.

However, existing forecasting systems struggle once predictions extend into the sub-seasonal range, roughly two to six weeks ahead. At this stage, forecasts focus more on probabilities, temperature anomalies, precipitation patterns, and an increased risk of certain weather conditions.

More accurate forecasts could have significant benefits, particularly for agriculture. Farmers could better prepare for periods of drought or excessive rainfall and adjust their plans for sowing, irrigation, fertilization, and harvesting accordingly. These forecasts are also important for emergency management and disaster preparedness.

Researchers at MIT are working to improve forecasting accuracy. Rather than simply training an AI model to make better predictions, they use AI to identify recurring or systematic errors in existing forecasting models that their developers have not previously recognized, then correct them. The study was published in Nature.

After applying their machine learning model to the gold standard among existing forecasting models, the researchers were able to double forecast reliability for the two-to-six-week period.

They also applied their AI to less advanced forecasting models, significantly narrowing the performance gap between those systems and the leading model. The MIT researchers hope to deploy the technique worldwide, particularly in regions where early warnings of natural disasters are especially important.