Scaling Hydrologic Models from Watersheds to River Basins Using Machine Learning and Coprocessors
Authors
Ethan T. Coon1* (coonet@ornl.gov), Soumendra Bhanja1, Sudershan Gangrade1, Bilal Iftikhar1, Julien Loiseau2, Dan Lu1, David Moulton2, Scott Painter1, Carl Steefel3
Institutions
1Oak Ridge National Laboratory, Oak Ridge, TN; 2Los Alamos National Laboratory, Los Alamos, NM; 3Lawrence Berkeley National Laboratory, Berkeley, CA
URLs
Abstract
Threats to water quality and quantity are of utmost concern as society grapples with a changing Earth system. Dwindling freshwater supplies in arid systems; nitrogen loading in agricultural systems; and pollutants, salinity intrusion, and flooding in coastal urban systems are of crucial importance in a changing climate. While many ESS projects continue to advance process understanding at scales at which those processes can be observed directly and modeled (often at the hillslope or small catchment scales), Earth and environmental system modeling projects face the difficult challenge of integrating process understanding into models at the global scale. To address this challenge, the ExaSheds project has developed and demonstrated a key linking capability that scales process understanding from the site level to full river basins.
The research team describes a model-centric scaling strategy—developed and partially demonstrated in the ExaSheds project—that combines machine learning and process-based simulation to enable high-resolution, process-based simulations of hydro-biogeochemical function at full river basin scales. The Advanced Terrestrial Simulator code (ATS) demonstrates how programming models that abstract the supercomputer’s architecture can enable high-resolution process-rich models informed by diverse data products to run on heterogeneous architectures at increasingly large scales. Machine learning is used to represent components of the hydrologic system that have uncertain process representation. For example, machine learning can be used to represent reservoir operations within an ATS model, focusing on the impacts of reservoirs on the river basin water cycle.
Finally, the Gunnison River Basin was used to demonstrate an approach where ML-based surrogate models are trained on ATS simulations of selected subbasins and then transferred spatially and temporally to predict long-time function of full river basins.