AI is making strides in weather forecasting, with models that outperform traditional systems in speed and resource efficiency. However, they face limitations; AI predictions are based solely on historical data, which becomes problematic during extreme weather events that haven’t been previously encountered.
A study from several universities reveals that while AI is effective for common weather forecasts, it tends to underestimate the severity of rare events, such as Category 5 hurricanes. For instance, when trained on data excluding strong hurricanes, models predicted a Category 2 hurricane instead of accurately forecasting a more severe storm, highlighting the risks associated with false negatives.
Traditional weather models rely on physics-based equations, allowing them to account for atmospheric dynamics, unlike AI models, which act like advanced autocomplete systems. The researchers found that AI could make better predictions if it had encountered similar extreme events in different regions.
To improve AI forecasting, blending traditional physics with AI techniques, like active learning to simulate rare weather events, is recommended. The goal is to enhance AI’s understanding of atmospheric dynamics for better predictions, especially regarding extreme weather. The study emphasizes the importance of recognizing AI’s limits while striving for improvements.
Source link
