Assistant Professor Yiqun Xie and coauthors received the Best Application Paper Award at the SIAM International Conference on Data Mining (SDM) on April 27 to 29, 2023, for their paper titled “Physics-Guided Meta-Learning Method in Baseflow Prediction over Large Regions” (Authors: Shengyu Chen, Yiqun Xie, Xiang Li, Xu Liang, and Xiaowei Jia).*
Deep learning has achieved promising success in computer vision and natural language processing tasks. Recent large language models such as ChatGPT also gained a tremendous amount of attention with their versatility and efficiency in performing human tasks. However, the success is still very limited in scientific domains where data are scarce and hard to collect.
This paper tackles the problem of baseflow prediction. Baseflow is the portion of the stream flow that is sustained between rainfall events and during dry periods, which is essential for ecosystem functioning. Due to the difficulty in collecting real observations, various physics-based models have been developed for baseflow estimation under different assumptions. In practice, however, it is often challenging to know which assumption – or if any existing assumption – aligns with the real complex environmental condition.
This work develops a physics-guided meta-learning framework to incorporate diverse knowledge from multiple physics-based models to enhance the training of the deep learning model, using limited observations. In particular, the approach learns to select a mixture of physics-based models to guide the training for different environment conditions. Experiment results in 60 river basins showed superior performance of the new approach compared to the other methods.
The paper shows the exciting potential of meta-learning-based integration between data-driven and physical models. The team will continue to explore new techniques along this direction through their collaborative projects from NSF and NASA.
Photo: Yiqun Xie. Courtesy of Xie.