Towards Hybrid Physics-Machine Learning Parameterizations - Employing Data Assimilation for Online Learning of Turbulence and Convection Closures in a Unified Scheme

Aug 5, 2026

Upload by Asher

研究聚焦于将物理约束的混合建模与机器学习参数化结合，用数据同化方式实现湍流与对流闭合的在线学习。文中提出在统一框架下，将物理模型结构、机器学习可校准的表示与观测/同化流程耦合，从而在动态环境中持续更新闭合关系并提升预测一致性。研究目标是解决传统静态参数化难以适配流体系统时变变化的问题，增强气候与大气数值模型的可用性与泛化能力。

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages99

Reading Time4 h 9 min

AccessFree to download

Acknowledgements

Advisor and committee contributions

EDMF and CliMA team support

Professional mentorship and personal thanks

Towards Hybrid Physics-Machine Learning Parameterizations: Employing Data Assimilation for Online Learning of Turbulence and Convection Closures in a Unified Scheme Thesis by Costa Christopoulos In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy CALIFORNIA INSTITUTE OF TECHNOLOGY Pasadena, California 2025 Defended November 8, 2024 ii © 2025 Costa Christopoulos ORCID: 0000-0002-8552-465X All rights reserved iii ACKNOWLEDGEMENTS Endless pages would not suffice to thank everyone who made my PhD years the happiest and most productive of my life. This experience surpassed even the highest expectations I had when I started applying to PhD programs in 2018, largely because of the people who surrounded me and the work I was doing. First, to my advisor Tapio Schneider. Above all else, Tapio was extremely patient and trusting. He had a lot of confidence in me—sometimes more than I had in myself—which boosted my self-assurance at the times I needed it most. The same goes for the rest of his students. When problems arose, he wanted to face them head-on, and not kick the can down the road. Having spent a career in climate modeling, he knows that unresolved issues reemerge and plague climate models later in complicated ways. In a world where it is tempting to quickly hack together code and stitch together limited solutions for flashy results, Tapio taught me the importance of digging into problems and addressing them one by one. For tackling any large-scale technological effort, like climate modeling, this is essential. This approach is, and will continue to be, the key to CliMA’s success. I also want to thank him for being patient with me throughout my journey, especially when code was not working or when I asked rudimentary questions. Tapio always tried to put the happiness and well-being of his students above all else. I’m also grateful to Tom Beucler, who, despite living in Switzerland, devoted his personal evening hours to meet with me on a weekly basis, for years. His patience and dedication to students are truly exceptional. To Christian Frankenberg, who served as an advisor for my second quals project and an unofficial mentor during my first two years. He guided me through the complexities of satellite inversionsand radiation codes, and our shared sense of humor made the journey all the more enjoyable. From Tapio, Christian, and Tom, who served as advisors (but not all on paper) during my time here, I learned distinct and invaluable lessons and approaches. I want to acknowledge Ignacio Lopez-Gomez, who initially took on the challenge of calibrating single-column models and taught me much of what I know today about working with the EDMF. To the entire early EDMF team—Yair Cohen, Charlie Kawczynski, Anna Jaruga, and Haaakon Ervik—working with you was a foundational experience, and this work would not be possible without each and everyone of you. To the restofthe CliMAteam, your advice and good company helped me iv tackle many challenges, whether by finding solutions through our conversations or simply by lifting my spirits. Even during tough times and tight deadlines at CliMA, laughter often echoed from the conference room, in the kitchen, and around the patio lunch tables. A special mention goes to Zhaoyi Shen for her endless support and deep, broad, and practical knowledge of climate modeling, fluid dynamics, numerics, and everything inbetween. There is perhaps no person who received more questions from me, and she (almost) always had the answer. And Ollie Dunbar, who patiently taught me and other students how to apply inverse methods to practical problems, always ready to help with issues and answer questions, big and small. To Daniel Rothenberg, my former manager and professional mentor. I looked up to him immensely and continue to. I joined an early-stage startup because of him, and stayed for the same reason. Together we faced seemingly innumerable challeng

FAQs

本文研究的核心问题是什么？

如何在统一框架下，用混合物理-机器学习参数化并结合数据同化，实现湍流与对流闭合的在线学习。

“在线学习”和“统一方案”在研究目标中扮演什么角色？

在线学习强调闭合关系随信息更新持续调整；统一方案则用于把物理约束、机器学习与同化流程组织在同一体系中。

研究主要面向哪些物理环节进行建模与改进？

湍流与对流的闭合（convection closures），通过在线更新以提升数值预测的适配性。

Aug 5, 2026

Upload by Asher