Process guided machine learning for water temperature prediction

Water temperature plays a key role in many societally relevant outcomes from thermonuclear cooling capacity to fish habitat to harmful algal bloom formation; delivering accurate temperature estimates is key to understanding the economic, social, and environmental impacts of changing water temperature in the United States. Many water temperature processes – stream-reservoir interactions, groundwater influence – are localized and difficult to represent in broad-scale process models or lack sufficient data for calibration, and therefore limit the prediction accuracy of process-based models.

This project will support methods development for combining advanced artificial intelligence models with process-based models of water temperature, both of which have significant drawbacks when used alone but have the potential to increase predictive accuracy when used together. Hybrid approaches are more flexible than process-based models and may improve prediction when certain processes are poorly characterized (e.g., groundwater). The new methods, tools, and code used to apply these techniques will be publicly available and applicable to other water resources research questions. For example, hybrid modeling has the potential to be useful for prediction of water characteristics where we have some understanding of the underlying processes that can be used to constrain machine learning models.