First-of-its-kind AI tool can predict water quality across the U.S.

Every day, the United States’ extensive water supply system faces pressure to deliver safe water. Now, University of Vermont (UVM) scientists have invented a new tool using AI to help communities better predict threats to their supply. 

New research published today by Vermont scientists shows how an already-existing computer system—the federal government’s National Water Model—can be modified, with AI and real-time data from sensors, to go beyond simply forecasting stream flow—to predicting water quality too. 

“This new tool can be implemented across the country and broadly utilized by folks that could use water quality forecasts in any number of applications,” said UVM’s Andrew Schroth, the study’s lead researcher. “With the first ever application of the National Water Model to predict water quality, we've opened a new window that can really benefit the country as a whole moving forward.” 

To test the tool in real-world conditions, the researchers focused on New York City’s water supply, an ideal testing ground due to the city's extensive network of sensors that monitor water flow and sediment—and the episodic nature of the problem that the NYC water supply faces. 

New York City’s Water Supply 

Publishing their findings in the Journal of the American Water Resources Association, the team tested the new tool in New York State’s Esopus Creek catchment in the Catskill Mountains. This waterway drains into the Ashokan Reservoir, which supplies approximately 40% of New York City’s daily drinking water and is part of the largest unfiltered water supply network in the country.  

A key concern for reservoir water quality is turbidity, a measure of water clarity affected by sediment and other material in the water column. When certain levels are exceeded, the NYC DEP must limit the supply from that reservoir, impacting the management and operations of the entire downstream network. The capacity to forecast high turbidity threats is critical for streamlining water supply operations. 

“When too much sediment comes into the reservoir during or after big storms, New York City has to limit supply and modify their operations,” says Schroth, a Research Associate Professor in the University of Vermont’s Department of Geography and Geosciences. He notes that Esopus Creek is prone to high turbidity due to high amounts of fine grain sediment from the glacial clays, silts, and gravels in the valley. When storms occur, stream banks erode, cutting into the glacial sediment and creating an elevated cloudiness which may linger for months, complicating forecasts and reservoir management. 

National Applicability 

Using New York City as a test case, researchers supported by the Cooperative Institute for Research to Operations in Hydrology (CIROH), a partnership based at the University of Alabama and supported by the National Oceanic and Atmospheric Administration and the U.S. Geological Survey, are working to apply the National Water Model beyond hydrologic forecasting.  

According to the National Weather Service, “the National Water Model is a hydrologic modeling framework that simulates observed and forecast streamflow over the continental United States.” The National Water Model allows anyone to access the data, giving users the ability to see how streams or creeks are impacted in the event of heavy rainfall, and if flooding can be expected. CIROH researchers nationwide are advancing innovative applications of the model to improve awareness of water availability and threats.  

The newly published research, led by UVM scientist, Dr. Andrew Schroth, and Utah State University scientist, Dr. John Kemper, in collaboration with UVM engineer, Dr. Kristen Underwood, NYC Department of Environmental Protection (DEP) geologist, Dany Davis, and U.S. Geological Survey scientists, focuses on using AI to combine the extensive datasets from the National Water Model with sensor datasets that collect data from streams in frequent intervals. By leveraging the relationship between the flow of water and concentration of sediment, the team was able to develop a robust forecast for water quality measurements.  

Across the United States, this tool can have widespread applications, allowing locations that typically face various water quality issues to better predict threats. If a water treatment plant has been implementing water quality monitoring, they now have the potential for real time capacity to understand how an upcoming storm will affect the water quality, giving them greater predictability in plant operation closures. Likewise, if an algal bloom is predicted with a storm onset, authorities will be aware of this water quality threat, closing down beaches for public health concerns. From an agricultural standpoint, farmers would know how much water is anticipated and what would be wet, guiding their use of fertilizer application. 

Kemper, who hails from Maryland and recently completed a post-doctoral fellowship at the University of Vermont before a move to Utah, acknowledges the impact of the work. “Turning a streamflow forecasting tool into a water quality forecasting tool paves the way for increasingly available forecasts to serve community needs”, creating water quality forecasts for communities and informing similar strategies for managing turbidity in basins worldwide.  

The model will set a precedent for studying other water quality components, extending far beyond New York City. The framework developed by the team has national applicability, as it can be adapted nationwide, allowing water plant operators and other managers to learn about their water quality constituent of interest, such as phosphorus, nitrogen, nitrate, turbidity, or chloride. This new model can transform how water quality can be anticipated and managed across the nation.