15vip太阳集团(中国)有限公司

“争流”学术沙龙（第十二期）：融合物理模型、机器学习与大样本观测的大尺度陆面水文模拟和研究

发布日期： 2024- 09- 01

访问次数:

报告时间：2024年9月3日15:00

报告地点：老图书馆4层4023会议室

报告题目：融合物理模型、机器学习与水文大数据的大尺度陆面水文过程模拟和研究

Unifying Physical Models and Machine Learning to Harness Large Hydrologic Observations and Characterize Terrestrial Water Cycle

报告人：冯大鹏博士

冯大鹏，斯坦福大学地球系统科学系博士后，斯坦福以人为本智能研究院（Human-Centered AI）Postdoctoral Fellow。2015年本科毕业于武汉大学，2018年硕士毕业于北京大学，2023年博士毕业于美国宾州州立大学。2023年至今在斯坦福大学从事博士后工作，研究主要包括：（1）系统性结合物理模型、深度学习和大量多源观测的陆面水文循环模拟与研究、大尺度水文模型和预报、植被-水关系及对气候响应等；（2）开发了能深度融合机器学习和物理过程的可微分水文模型(differentiable hydrologic modeling framework)。相关研究成果发表在《Nature Reviews Earth & Environment》、《Nature Communications》、《Geophysical Research Letters》、《Water Resources Research》、《Geoscientific Model Development》等期刊。

报告摘要：

The eruption of earth observations from both ground and satellite sources provides unprecedented opportunities for hydrologic sciences. Traditional physical models are valuable tools for hydrologic simulations, but are hindered by limitations to fully embrace large observations. Pure machine learning methods can automatically learn from big data for superior performance, while they still lack interpretability to freely answer scientific questions. We desire next-generation models that can harness the strengths of both pathways. This talk will introduce the developed differentiable hydrologic modeling framework which is a generic way to integrate physical models and neural networks. This differentiable framework can automatically learn from large observations like machine learning models, while still preserving internal physical clarity to accurately simulate a full set of variables in terrestrial water cycle. Neural network components can be flexibly embedded to support data assimilation, develop automatic parameterizations, and learn unclear hydrologic functions. This talk will showcase two applications of this hybrid modeling framework for large-scale simulations of rainfall-runoff processes and the soil-plant-atmosphere continuum, constrained by ground and remote sensing observations, respectively. Our studies demonstrate that differentiable modeling can not only achieve state-of-the-art hydrologic prediction performance in both data-rich and data-scarce scenarios, but also provides promising opportunities to improve model structure deficiencies, identify unclear hydrologic relationships and facilitate knowledge discovery.