Collaborative Research: III: Small: Physics Guided Graph Networks for Modeling Water Dynamics in Freshwater Ecosystems

点击翻译按钮获取中文摘要

Fresh water plays an important role for the global economic, food, water, and energy networks, but freshwater ecosystems continue to degrade due to pressures from increasing demands for freshwater ecosystem services and a shifting climate. Timely monitoring of water properties can provide useful information for sound policy and management decisions to address important water-related challenges such as droughts, floods, and water security. Moreover, the information of water properties such as water temperature and streamflow can help better understand relevant biogeochemical and ecological processes in the water cycle. The recent investment on large-scale water data repositories provides a tremendous opportunity for using machine learning to capture complex water dynamics over space and time. In particular, graph neural networks have shown great promise for modeling interactions amongst streams in large river basins. However, in the absence of underlying physical knowledge, direct applications of existing graph-based models remain limited in capturing complex water-related processes, modeling the shift of data distribution caused by human infrastructure or changing climate, and learning from a paucity of data samples. To overcome these limitations, this project will explore a deep coupling of graph network models with physical knowledge to model complex, non-stationary, poorly observed water dynamics in freshwater ecosystems. This project will provide research opportunities to graduate and undergraduate students from diverse backgrounds, and the results of this project will be incorporated into curriculum development. This project aims to develop new physics-guided graph network models by designing new model architectures, learning strategies, and initialization methods. This project will also explore different ways to leverage physical knowledge, both directly by integrating physics from known mathematical equations, and indirectly by making use of the knowledge embodied in existing physics-based models. In particular, there are three innovations that are pursued in this project. First, new graph-based architectures will be developed to model the complex nature of physical objects and the dynamic interactions between physical processes. Second, new graph-based continual learning strategies will be investigated to model long term system evolution caused by newly added infrastructure and changing climate. Third, new model initialization methods will be developed by transferring knowledge from existing physics-based models to the proposed graph network models to facilitate learning physically consistent patterns in data-scarce scenarios.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.