New AI Tool Boosts Extreme Weather Forecasts

Extreme weather events pose significant threats, but improved forecasting offers a crucial advantage. A new post-processing tool, detailed in the journal Nature, utilizes machine learning to enhance the accuracy of subseasonal weather outlooks, extending predictions from two to six weeks.

Developed by Judah Cohen and Lester Mackey, this Probabilistic Bias Correction (PBC) tool is model-agnostic, meaning it can correct errors in both traditional numerical and newer AI weather models. In a recent forecasting competition, it produced the most accurate global forecasts across all variables and periods.

The PBC tool has demonstrated remarkable improvements. It doubled the subseasonal accuracy of leading AI systems and enhanced numerical models for 91% of pressure, 92% of temperature, and 98% of precipitation targets. Its most impactful development is the significant improvement in predicting extreme weather events.

Forecasting extreme heat, precipitation, and high sea level pressure three weeks in advance saw substantial gains. Researchers also found that PBC improved flood prediction skill by 76% to 92% for real-world floods over the past decade. For severe floods, forecasting skill increased by 86% to 150%.

This open-source tool, which does not require supercomputers, aims to provide more actionable extreme weather forecasts globally. The meteorological office of the Republic of Kenya is already using it to refine monsoon forecasting, highlighting its accessibility and potential to save lives and reduce damages as climate change intensifies extreme weather.