The Upscaling Problem: Diagnosing Forecast Error Growth in Next-Generation AI and Physics-Based Weather Prediction Models

Descripción del proyecto

Modelos de predicción meteorológica basados en inteligencia artificial para patrones de circulación complejos

La predicción meteorológica ha mejorado gracias a los avances en inteligencia artificial (IA) y a los modelos físicos de resolución kilométrica, capaces de representar la convección profunda. En muchos casos, estos nuevos sistemas superan a los modelos globales actuales y ofrecen previsiones más fiables. Sin embargo, a pesar de estos avances tecnológicos, los sistemas basados en IA pierden precisión ante patrones complejos de circulación, acoplados a diferentes escalas y asociados con frecuencia a fenómenos meteorológicos extremos. En el proyecto UPDAP, que cuenta con el respaldo de las acciones Marie Skłodowska-Curie, se llevará a cabo la primera comparación sistemática entre modelos físicos y modelos de IA. El análisis permitirá determinar en qué medida reproducen el acoplamiento entre la convección y la corriente en chorro y cómo afecta este proceso a la capacidad de predicción en Europa.

Objetivo

Operational weather prediction is entering a hybrid era, combining rapidly developing global AI-based models with a new generation of kilometer-scale physics-based models that explicitly resolve deep convection. AI-based systems often outperform current global models that parameterize convection on standard forecast skill metrics, while kilometer-scale models promise improved predictability through more realistic representations of mesoscale processes. However, major uncertainties remain regarding how these emerging paradigms simulate and predict complex, scale-coupled circulation patterns. A critical example is the interaction between organized mesoscale convection over North America and the jet stream. In contemporary physics-based models, misrepresentation of this coupling is a well-known driver of sudden degradation in forecast skill over Europe, often coinciding with high-impact weather events. Recent evidence indicates that global AI-based models are also vulnerable to sharp drops in European forecast skill when circulation patterns favor convection–jet stream coupling. Thus, both AI-based and physics-based approaches introduce significant uncertainty for forecast users, particularly during high-impact weather events. This makes it essential to conduct process-level evaluations across modelling paradigms to uncover the mechanisms within these circulation patterns that trigger rapid forecast error growth. UPDAP will provide the first systematic intercomparison of AI-based and physics-based models (including kilometer-scale models) to determine how well they capture convection–jet stream coupling and its influence on European predictability. The project’s outcomes will deliver actionable insight forecast data users, guiding the safe and reliable integration of AI and kilometer-scale physics models into future operations.