AI Models Struggle with Extreme Weather Events

Artificial intelligence has been showing positive results with its ability to forecast weather days in advance. AI models can often run faster and use far less computing power compared to traditional weather forecasting systems. However, AI models struggle when it comes to extreme weather events due to its ability to only predict similar events to what has already happened in the past.

A research team from the University of Chicago, New York University, and the University of California Santa Cruz has begun to study this issue. They set out to determine if a dangerous weather event hasn’t happened in the training data, will the AI model still predict it? They tested this theory by training an AI weather model with data that excluded all hurricanes stronger than a category 2.

When they asked the model to forecast a scenario that would typically lead to a category 5 hurricane, the AI model consistently underestimated the event. The model always predicted that it would only be a category 2 hurricane. In weather forecasting, largely underestimating an event like that could lead to costly and dangerous impacts.

In similar fashion, AI models have been used to try and predict the long-term risks of a warming climate and future climate scenarios. But if AI struggles to forecast extreme events that haven’t happened before, its usefulness for those tasks becomes limited.

Researchers believe the solution to this problem is to blend traditional physics models with AI through active learning, in hopes that the AI models can learn atmospheric dynamics and figure out how to forecast the more rare and extreme weather events.