报告题目：Conservation principles for data assimilation and neural networks
报告专家：Dr. Yvonne Ruckstuhl
报告时间：2023年5月10日（周三）16:00
会议形式：ZOOM-线上会议（会议ID: 329 593 0089, 会议密码：87654321）
会议链接(无需密码)：https://us02web.zoom.us/j/3295930089?pwd=Tis2eU4vVE5mZmI4NGxpVXRDRVhxQT09
主持人：曾跃飞 教授
专家简介：
Dr. Yvonne Ruckstuhl, bachelor and master in applied mathematics at Delft University of Technology, Netherlands; PhD in meteorological institute at the University of Munich (LMU). Since 2019 she has been working as Postdoc at LMU. She is interested in advanced data assimilation algorithms, model uncertainties and machine learning, and has published more than 10 SCI articles. Her recent work in machine learning has been selected as highlight by the Nonlinear Processes in Geophysics.
报告摘要：
Numerical discretization schemes are often designed to preserve the most important conservation properties of the continuum system, like mass conservation. However, data assimilation algorithms and neural networks typically do not preserve physical properties, leading to biases. In this talk, we argue that data assimilation algorithms and neural networks should incorporate some of the conservation principles, following the similar principle of design as NWP models. In particular we show in various configurations that forecast errors are reduced significantly when physical constraints are included in the data assimilation. We also show that neural networks used to reduce model error due to unresolved scales perform better when mass errors are reduced.
欢迎广大师生踊跃参加！
气象灾害教育部重点实验室
气象灾害预报预警与评估协同创新中心
大气科学学院
2023.05.04
