MODULAR COMPOSITIONAL LEARNING FOR WATER TEMPERATURE MODELING V2.0: MORE LAKES, RESTRAINTS, AND MEMORY

One-dimensional water temperature models provide a valuable tool to quantify future climate change impacts on lakes and reservoirs at the global scale. Recently, the novel paradigm of Knowledge-Guided Machine Learning (KGML) has highlighted the increased performance of 1D water temperature projections by combining process-based modeling with deep learning. Modular compositional learning (MCL), a design based on the KGML paradigm, highlighted that a deep learning model embedded in a process-based model framework can improve predictions of physical limnological variables with a high signal-to-noise ratio while limiting physically unrealistic density violations. In the hybrid MCL model, turbulent diffusive transport is simulated through the deep learning model to decrease the uncertainty of empirical process parameterizations. Here, we improve upon the current hybrid MCL model by changing its training algorithm from sequential to recurrent. This update allows the hybrid MCL model to utilize past information for future predictions. Further, we incorporate a loss function to penalize energy conservation violations. This loss function helps to prevent physically unrealistic results by the deep learning model. Finally, we train the deep learning model for turbulent diffusive transport on a set of lakes from Denmark, Sweden, Switzerland, Germany, and the US to increase its generalizability. The modified hybrid MCL model produces more robust results than its previous version with improved water temperature projections across a variety of lake types, highlighting the utility of this approach for hydrodynamic modeling.

Primary Presenter: Robert Ladwig, Aarhus University (rladwig@ecos.au.dk)

Authors:

Robert Ladwig, Aarhus University  (rladwig@ecos.au.dk)

Abhilash Neog, Virginia Tech  (abhilash22@vt.edu)

Arka Daw, Virginia Tech  (dawa@ornl.gov)

Cayelan Carey, Virginia Tech  (cayelan@vt.edu)

Paul Hanson, University of Wisconsin-Madison  (pchanson@wisc.edu)

Anuj Karpatne, Virginia Tech  (karpatne@vt.edu)