“Our neural network approaches calibration by learning from the large datasets we have from past measurements, while also taking into account physics-based information from the NWM,” Song said. “This allows large datasets to be processed very efficiently with a higher level of consistency and reliability without losing the level of detail that physics-based models provide.”
Shen said this approach to calibration is not only efficient, but very consistent, regardless of the region being simulated.
“The old approach is not only very inefficient, but it's completely contradictory,” Shen said. “With a new approach, you can use the same process to create simulations regardless of the area you are trying to simulate. As you process more data and create more predictions, the neural networks continue to improve.
According to Shen, their model is a candidate for use in the NWM's next-generation framework, which NOAA is developing to improve national flood forecasting standards. Although not yet selected, Shen said the model is “very competitive” as it is already tied to this operational framework. However, while it may take some time for model users to become familiar with the AI components of the model, Shen explained that careful independent evaluation is necessary to demonstrate that model accuracy can be trusted even in untrained scenarios. The team is working to close the final gap – improving the model's prediction capabilities from daily to hourly – more useful for operational applications such as hourly flood clocks and warnings. Shen praised the research and manipulation research work to Leo Lonzarich, a doctoral candidate in civil engineering, noting that developing a framework that other researchers can expand is the key to solving the problem and evolving the model as a community.
“When a model is trained, predictions can be generated at an unprecedented rate,” Shen explained. “In the past, it took several weeks to generate 40 years of high resolution data through NWM, and many different supercomputers had to work together. Now, one system can do that in just hours.
Although these models are primarily used for flood prediction, simulations provide hydrologists with information that can be used to predict other major events such as droughts. Such forecasts can be used to inform water resource management. This could have an impact on agriculture and sustainability research, Shen said.
“Our model is physically interpretable, allowing us to explain the characteristics of river basins, such as soil moisture, basic flow velocity of rivers, and groundwater charging. “We can better understand ecosystems and the natural systems that play an important role in supporting ecosystems and the ecosystems within them.”
Alongside Shen and Son, co-authors of the Pennsylvania paper include TADD Bindas, who recently completed his PhD in Civil and Environmental Engineering. Kathryn Lawson, researcher in the study of deep learning in hydrology. Doctoral candidates in Private and Environmental Engineering in Pennsylvania Haoyu Ji, Leo Lonzarich, Jiangtao Liu, Farshid Rahmani, Kamlesh Arun Sawadekar.
Additional co-authors include Wouter JM Knoben Cyril Thébault and Martyn P. Clark of the University of Calgary. Katie Van Werkhoven, Sam Lamont and Matthew Denno, Research Triangle Institute; Mann Pan and Yang Yang of the University of California, San Diego, Oceanography Institute, San Diego. Michigan State University Jeremy Rapp. Mukesh Kumar, Richard Adkins, James Halgren, Trupesh Patel, Arpita Patel, University of Alabama. The Pennsylvania team thanked the University of Alabama for the computing support they provide.
The National Oceanic and Atmospheric Administration of Hydrologies Research Institute for Operations in the Cooperative Agreement, as well as the U.S. Department of Energy, the National Center for Energy Research and Science Computing, and the California Department of Water Resources Atmospheric River Program supported this research.
In Pennsylvania, researchers solve real problems that affect the health, safety and quality of life of people across the federal, nation, and around the world.
For decades, federal support for research has encouraged innovations that make our country safer, make our industry more competitive, and strengthen our economy. Recent federal funding has threatened this progress.
Learn more about what it means for federal funding cuts to our future Research or regression.
