Comparison Of Physics-Informed Mass-Conserving Perceptron Against Data-Driven Neural Network And Physical-Conceptual Models In Modeling The Hydrologic Systems
Presenter: Yuan-Heng Wang1
Co-Author(s): -
Advisor(s): Dr. Hoshin Gupta
1Department of Hydrology and Atmospheric Sciences, University of Arizona
Generic Gated Recurrent Neural Network (particularly Long Short-Term Memory Networks) are currently the most accurate and extrapolatable predictive models available for time-series prediction for earth and environmental science problems. However, the relatively poor physical explainability of such ML-based approaches impedes widespread acceptance of DL-based models by the hydrological science community. Therefore, I will introduce our development of a Mass-Conserving Perceptron (MCP) structure for addressing the “accuracy-interpretability” dilemma in the context of physics-informed machine learning. I will demonstrate how the inherent isomorphism between recurrent neural network structures and dynamical systems enables the representation of important hydrological process (mass/energy-conserving) in a meaningful manner, thereby bridging the gap between current physics-based modeling and ML learning theory. Several MCP-based structures are tested in the context of daily rainfall-runoff modeling for the Leaf River catchment. Particular attention is given to exploiting ML technologies that 1) ensure robust training to achieve good generalization performance, and 2) benchmark the performance against pure data-driven neural network as well as physical-conceptual models. Overall, this study provides the insight regarding how to optimize the information extraction through the management of the complexity of the system architecture for the development of synergistic physics-AI modeling approaches.