Researchers from Princeton will help lead a $1 million project funded by the National Science Foundation (NSF) that will use artificial intelligence to simulate the nation’s natural groundwater system in an effort to improve water management and help people better prepare for flooding and drought.
The project was one of 29 nationwide that received a total of $27 million for the first phase of the NSF’s Convergence Accelerator program. Now in its second year, the initiative is designed to advance team-based multidisciplinary research that address challenges of national importance and that will produce societally valuable results in the near future.
The research team is led by Laura Condon, assistant professor of hydrology and atmospheric sciences at the University of Arizona (UA), and includes Princeton co-principal investigators Reed Maxwell, professor of civil and environmental engineering and the Princeton Environmental Institute (PEI), and Peter Melchior, assistant professor of astrophysical sciences jointly appointed in Princeton’s Center for Statistics and Machine Learning. Project co-leads also include Patrick O’Leary, assistant director of scientific computing at scientific software company Kitware, and Nirav Merchant, director of the UA Data Science Institute and co-head of CyVerse, a national computational infrastructure for the life-sciences that is funded by NSF.
The project will combine the researchers’ strengths in data science, machine learning and hydrology — which examines the dynamics and management of Earth’s water cycle — to improve hydrologic forecasting, or the prediction of how much groundwater is available, how it can be sustainably managed, and how it will influence the severity of extreme events.
Groundwater can prevent drought or exacerbate flooding, but predicting how these events could be affected requires knowing how much groundwater there is, said Maxwell, who will head the Princeton end of the project, which will be based in PEI. Currently, effective and comprehensive hydrologic forecasting is hindered by a “patchwork” of models and data sources that are maintained by various scientists and institutions, he said.
“We don’t know how much groundwater we have, so we don’t know how much we can rely on it during normal years — let alone drought years — nor the extent to which it could exacerbate flooding, especially in mountain systems,” Maxwell said.
“This project brings many data sets and model results together in a seamless framework,” Maxwell said. “This is a complicated problem that bridges disciplinary data, complex numerical simulations, substantial software development, data science and machine learning, user engagement, and education and outreach. We are bringing all of those elements together in a coherent way, which I believe is very unusual — if not unique — for an NSF-funded research project.”
Groundwater models normally require more computer capacity than most researchers have at their disposal, Condon said. The Convergence Accelerator project builds on a national-scale hydrologic modeling platform Condon and Maxwell work on known as HydroFrame — which allows for the rapid and accessible modeling of any watershed in the United States — by fine-tuning that framework to address specific water-management needs. The U.S. Bureau of Reclamation is partnered with the team to implement their data in managing water in the western United States. The team also will work with the National Oceanic and Atmospheric Association’s Geophysical Fluid Dynamics Laboratory — located on Princeton’s Forrestal Campus — to integrate their simulations of hydrological systems into global climate models.
“Simulating groundwater flow and groundwater surface water interactions is very computationally challenging,” Condon said. “Our team has developed some of the first national-scale groundwater simulations, but they require millions of core hours on super computers to generate, which can be a significant barrier for water managers and planners.
“Our project will leverage these computationally intensive and scientifically rigorous simulations to build machine learning models tailored to water management questions that can be easily built and run,” Condon said.
Phase one of the Convergence Accelerator program provides funding for a nine-month period to develop the initial concept further, identify new team members, participate in an innovation curriculum, and develop an initial prototype. At the end of phase one, each team participates in a pitch competition and a proposal evaluation. Teams that proceed to phase two receive two-year grants of up to $5 million; phase two projects from the inaugural Convergence Accelerator awardees were announced Sept. 3.
During the first phase, the Princeton/UA project will build an educational curriculum in preparation for a phase-two mentorship program intended to increase the participation of underrepresented students in the sciences.
“Water is a great way to engage students,” Maxwell said. “Through this initiative, we will develop undergraduate-level projects designed to connect students directly to real-world challenges.”
The project “Hidden Water and Hydrologic Extremes: A Groundwater Data Platform for Machine Learning and Water Management” was funded by the National Science Foundation’s Convergence Accelerator (C-Accel) program.