Traditional models still outperform AI for extreme weather forecasts (www.carbonbrief.org)

A recent study published in Science Advances reveals that traditional physics-based weather models outperform AI models when predicting record-breaking extreme weather events. Researchers evaluated the accuracy of both approaches by simulating thousands of extreme temperature and wind occurrences from the years 2018 and 2020. The findings indicated that AI models not only underestimated the frequency and intensity of such events but also struggled particularly in forecasting unprecedented extremes. This highlights a significant caution for the AI/ML community against quickly replacing traditional forecasting methods with AI models, which are still evolving and proving insufficient for critical tasks such as early warning systems. The study emphasizes the limitations of AI models, which rely heavily on historical training data and are less effective at simulating new weather patterns, particularly during extreme conditions. The authors suggest that while AI technologies are becoming more integrated into weather forecasting, their current capabilities warrant continued reliance on traditional models for accuracy, especially as early forecasts are crucial for minimizing risks associated with severe weather. Future research directions, including the development of newer probabilistic models that better encapsulate extremes and a hybrid approach combining physics-based and AI methodologies, promise potential advancements in predictive accuracy and operational efficiency in meteorology.