BK21 FOUR

연세대학교 지구·대기·천문 교육연구단
INSTITUTE OF EARTH ATMOSPHERE ASTRONOMY

Earth & Universe 2050

세미나

home  >  BK광장  >  세미나
2024년 9월 24일(화) 세미나 안내
  • 관리자
  • |
  • 64
  • |
  • 2024-09-19 14:17:39

제목: 인공지능 기반 전지구 기후예측모델 개발 및 최신연구 동향

 

연사강대현 박사 (KIST 기후탄소순환연구단)

 

일시: 2024년 9월 24일 화요일 16:00

장소과학관 B102호

 

Abstract:

Recent deep learning-based models have shown great potential in predicting and interpreting weather and climate phenomena. For example, the data-driven weather forecast models (e.g., GraphCast and Pangu-Weather) showed better forecast skills of the global atmosphere within a week than the operational numerical weather prediction. However, with their short history, the understanding of trained physical processes in the deep learning-based model is not satisfactory, which limits its predictability until two weeks. Motivated by the above, this study aims to improve the accuracy and capability of data-driven global climate prediction at sub-seasonal timescales. This study uses a deep learning-based model trained with daily-mean atmospheric variables in the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis (ERA5) at 250-km horizontal resolution for 1979-2015. A method to better structuralize the horizontal structure of the global atmosphere into the deep learning model shows great potential to improve data-driven prediction skills, which exhibit reliable deterministic forecast skills within a week or longer prediction. The inference results also exhibit realistic sub-seasonal climate variability, such as eastward propagation of the MaddenJulian Oscillation (MJO), indicating that the realistic physical processes are adequately trained in the data-driven method. The results of this study shed light on the necessary processes in the model architecture for state-of-the-art global climate prediction.

 

이전글 2024년 9월 20일(금) 세미나 안내
다음글 2024년 9월 27일(금) 세미나 안내

비밀번호 입력
비밀번호
확인
비밀번호 입력
비밀번호
확인
TOP