-
- GC43F-0860: Forecasting the Future with Yesterday’s Climate: Temperature Bias in AI Climate and Weather Models
-
Board 0860‚ Hall EFG (Poster Hall)NOLA CC
Author(s):Generic 'disconnected' Message
Jacob Landsberg, Colorado State University (First Author, Presenting Author)
Elizabeth Barnes, Colorado State University
In recent years, a surge of data-driven AI climate and weather models has demonstrated predictive skill rivaling, and sometimes surpassing, that of traditional physics-based models. Unlike these traditional models, which rely on physical equations, AI models infer relationships based on historical datasets. One key concern is that long-term trends, such as global warming, may not be well captured if the training data is from a colder historic climate. In this study, we examine temperature biases in two leading AI models: Ai2 Climate Emulator version 2 (ACE2) and FourCastNet v2 small model. We find that ACE2 predicts temperatures too hot for the past (1940–1970) and too cold for recent decades (1970–2020), suggesting it may not fully capture the effects of a changing climate. Similarly, FourCastNet underpredicts temperatures for 2020–2025. This bias is strongest on the hottest days, which are underrepresented in the historic training data. Our results reveal a key limitation of today’s AI forecasting: difficulty adapting to a changing climate with limited modern data. Solving this will be essential for reliable predictions of extreme events in a warming world.
Scientific DisciplineSuggested ItinerariesNeighborhoodType
Enter Note
Go to previous page in this tab
Session
