Presentation Information
[U17-P06]Reproduction of South Coast Low-Pressure Systems Using Machine Learning Models
*Komori Yuichiro1, Kazuya Yamazaki2, Yuki H. Takano3, Hiroaki Miura1 (1.The University of Tokyo, 2.Information Technology Center, The University of Tokyo, 3.Meteorological Research Institute)
Keywords:
Meteorology,Machine Learning
South Coast low-pressure systems that pass along the southern coast of Japan frequently cause heavy snowfall over the Kanto region. Accurate prediction of these events remains challenging because snowfall distribution strongly depends not only on the synoptic-scale cyclone structure but also on mesoscale cold-air retention, near-surface temperature, and topographic effects. In this study, we evaluate the ability of state-of-the-art machine learning–based global weather prediction models to reproduce such South Coast snowfall events and investigate the atmospheric conditions favorable for their development.
We analyze representative snowfall cases associated with South Coast cyclones using three machine learning models: GraphCast, AIFS, and Pangu-Weather. Forecasts are initialized from reanalysis fields and integrated up to 120 hours. Model performance is assessed in terms of cyclone tracking, upper-level trough position, frontal structure, surface temperature, and precipitation distribution.
All models successfully reproduce the large-scale synoptic features of the events. In particular, the cyclone track along the southern coast of Honshu and the position of the upper-tropospheric trough are reasonably well captured. The evolution of sea-level pressure patterns and the overall frontal configuration are also consistent with reanalysis data. These results indicate that machine learning models effectively represent synoptic-scale baroclinic development.
However, substantial discrepancies are found in mesoscale and near-surface fields. The retention of cold air over the Kanto Plain, surface temperature distribution, and localized precipitation intensity show notable differences from reanalysis. In several cases, snowfall areas are underestimated or displaced. These results suggest that while synoptic-scale structures are well represented, mesoscale processes and topographically dependent quantities remain difficult for current machine learning models to reproduce.
To further investigate the environmental conditions conducive to South Coast snowfall events, sensitivity experiments are conducted. First, perturbations are added to the climatological mean field before running the machine learning models. In these experiments, cyclones tend to intensify east of Japan rather than along the southern coast, indicating that small deviations from the mean state are insufficient to produce typical South Coast cyclogenesis. Second, the composite mean field derived from multiple South Coast cyclone cases is used as the initial condition. In this setup, although cyclone intensity is generally weaker than in real events, the South Coast frontal structure is reproduced.
Our ultimate objective is to clarify what large-scale atmospheric configurations favor snowfall-producing South Coast cyclones and to identify how different machine learning models represent these processes. Comparative analysis reveals differences among models in cyclone intensity evolution, precipitation response, and sensitivity to initial perturbations. Understanding these characteristics will contribute to improving data-driven weather prediction and clarifying the dynamical requirements for heavy snowfall events in Kanto Region.
We analyze representative snowfall cases associated with South Coast cyclones using three machine learning models: GraphCast, AIFS, and Pangu-Weather. Forecasts are initialized from reanalysis fields and integrated up to 120 hours. Model performance is assessed in terms of cyclone tracking, upper-level trough position, frontal structure, surface temperature, and precipitation distribution.
All models successfully reproduce the large-scale synoptic features of the events. In particular, the cyclone track along the southern coast of Honshu and the position of the upper-tropospheric trough are reasonably well captured. The evolution of sea-level pressure patterns and the overall frontal configuration are also consistent with reanalysis data. These results indicate that machine learning models effectively represent synoptic-scale baroclinic development.
However, substantial discrepancies are found in mesoscale and near-surface fields. The retention of cold air over the Kanto Plain, surface temperature distribution, and localized precipitation intensity show notable differences from reanalysis. In several cases, snowfall areas are underestimated or displaced. These results suggest that while synoptic-scale structures are well represented, mesoscale processes and topographically dependent quantities remain difficult for current machine learning models to reproduce.
To further investigate the environmental conditions conducive to South Coast snowfall events, sensitivity experiments are conducted. First, perturbations are added to the climatological mean field before running the machine learning models. In these experiments, cyclones tend to intensify east of Japan rather than along the southern coast, indicating that small deviations from the mean state are insufficient to produce typical South Coast cyclogenesis. Second, the composite mean field derived from multiple South Coast cyclone cases is used as the initial condition. In this setup, although cyclone intensity is generally weaker than in real events, the South Coast frontal structure is reproduced.
Our ultimate objective is to clarify what large-scale atmospheric configurations favor snowfall-producing South Coast cyclones and to identify how different machine learning models represent these processes. Comparative analysis reveals differences among models in cyclone intensity evolution, precipitation response, and sensitivity to initial perturbations. Understanding these characteristics will contribute to improving data-driven weather prediction and clarifying the dynamical requirements for heavy snowfall events in Kanto Region.