How much fresh water is in the United States? It's a tough question, since most of the water is underground, accessible at varying depths. In previous decades, it's been answered indirectly from data on rainfall and evaporation. Knowing how much groundwater is available at specific locations is critical to meeting the challenges of water scarcity and contamination.
Now, researchers at Princeton University and the University of Arizona have answered this question in unprecedented breadth and detail. Combining direct measurements with artificial intelligence methods, their new map estimates groundwater depth across the continental United States at a resolution of around 30 meters (98 feet).
These detailed estimates led to a new answer: a total of 306,000 cubic kilometers of water, or more than 13 times the volume of all the Great Lakes combined. While this data-driven method shows an amount of water that's in line with earlier studies, it does reveal supplies of shallow groundwater that were previously unknown. The work provides a foundation for further research as well as local and regional decision-making around irrigation, conservation and water infrastructure.
"Given all the things we do know about the planet, we don't actually know how much water we have," said senior study author Reed Maxwell , the William and Edna Macaleer Professor of Engineering and Applied Science at Princeton University. "And since most of it's in groundwater, knowing how much surface water we have is only about 1% of the total. That's where this becomes a hard problem. That's where it becomes an interesting science problem."
Despite groundwater's crucial role in ecosystems and as a resource for drinking water, agriculture and industry, direct measurements of groundwater depth are sparse across many regions of the United States and the world. To create the map, Maxwell and his co-authors combined more than a million direct measurements of groundwater depth with regional climate and geological data. They used this data to train AI algorithms that estimated groundwater depth at sites where measurements were not available. Their results were published Jan. 14 in the journal Communications Earth & Environment.
In addition to calculating the overall volume of groundwater, they estimated the depth of the water table, or the depth below ground at which groundwater is available. They divided the continental United States into a grid of more than 8 billion squares, each measuring 30 meters on each side, and estimated groundwater depth for each square. Combining this depth information with estimates of rock and sediment porosity allowed the team to calculate groundwater volume.
The groundwater depth estimates relied on data from United States Geological Survey groundwater monitoring wells and from previous studies, with measurements ranging from the years 1895 to 2023. This broad time span was necessary because more than half of the sites used in the study were only measured at a single time point. Most of the data was collected after 1970, said Maxwell.
"We had to [combine data from different time periods] to have enough to use a purely data-driven approach. These problems need big data, and we needed a lot of observations to be able to have a reliable model," said Maxwell, who is also a professor of civil and environmental engineering and the High Meadows Environmental Institute . "We call it a modern estimate," rather than an estimate at a specific point in time, he explained. The map hints at impacts of groundwater pumping in agricultural regions, but more data would paint a clearer picture of this trend.
It's counterintuitive, said Maxwell, but this is the first large-scale, hyper-resolution groundwater model that uses actual observations of groundwater. Previous models have been mainly physics-based, meaning they rely on equations that represent the flow of groundwater under different conditions, and have mapped water table depth at a resolution of about 1 kilometer.
The team's data-driven, AI-based model achieves a spatial resolution more than 1,000 times greater than physics-based models. The new approach also uses far less computing power, said lead author Yueling Ma , a former postdoctoral researcher in Maxwell's group who is now a research fellow at Forschungszentrum Jülich in Germany.
Notably, the new model calculates the uncertainty of the groundwater depth estimate at a given location. This calculation relies on a machine-learning method called a random forest. Co-author Peter Melchior , assistant professor of astrophysical sciences and the Center for Statistics and Machine Learning at Princeton, explains: "For each location, the method uses 300 decision trees. Each of these trees is trained slightly differently and is trying to solve the same problem. So, they find different solutions to it. If you have a full forest of them, you can use the variation between them as an estimate of the underlying uncertainty."
Overall, the model's uncertainty is higher in the western United States than in the East, which has important implications because the West generally has deeper water tables and is more dependent on groundwater for irrigation and drinking water, said Melchior.
Since 2020, Melchior and Maxwell have helped lead the development of the HydroGEN (Hydrologic Scenario Generation) platform, a suite of digital tools for understanding U.S. hydrology, along with University of Arizona researchers Nirav Merchant and Laura Condon, the project's lead investigator. The effort has been funded by a $5-million grant from the National Science Foundation's Convergence Accelerator program, which aims to accelerate real-world applications of research through interdisciplinary approaches. Recently this project was awarded a $500,000 supplement from the NSF National AI Research Resource to further democratize its datasets and to accelerate machine learning nationally.
On the HydroGEN team, "we have hydrologists, we have machine-learning experts, we have software developers," said Condon , a co-author of the study and a professor of hydrology and atmospheric sciences at the University of Arizona. This collaboration "allows us to go from hydrology ideas to really building something that other people can access and see and use. A lot of the work we did through the Convergence Accelerator has been focused on user experience and software development."
The model's outputs are publicly available on the team's HydroFrame platform, which makes hydrology data and simulations broadly accessible to researchers and decision-makers alike.
"There's a wide range of people who need to understand how much [groundwater] there is, how deep it is, how accessible it is," said Maxwell. "And those are just the immediate management needs that'll be met by this."
Lead author Ma said that researchers focusing on geochemistry and water quality have approached her about using the team's data to guide their own modeling experiments.
On a practical level, said Maxwell, the new map's high resolution could be valuable for farmers or other local and regional decision-makers. Much of our agriculture depends on center-pivot irrigation, in which a large sprinkler attached to a single well provides water to crops over an area of 500 square meters (about one-eighth of an acre). There are more than 14 million center pivots in the Ogallala Aquifer of the Great Plains, for example.
"It's this local decision [of how to irrigate] that is made millions of times," said Maxwell. And his team's new model accounts for the shallow groundwater that's critical to agriculture in a way that previous models overlooked.
"Because we use machine learning, these new techniques, and high-resolution big data products, we can get at that piece that's never been never been able to be calculated or estimated before," he said.
In collaboration with local experts, the team is beginning to expand its method globally. Ma, now in Germany, and co-author Julian Koch of the Geological Survey of Denmark and Greenland are focusing on areas of Europe. Maxwell recently returned to Princeton from a sabbatical in Australia, where he is working with hydrologists to build both physics-based and machine-learning groundwater models for the continent. Others in his group are leading similar work in Brazil. He noted that the approaches developed for the U.S.-focused map can be adapted to model groundwater in regions with less available data.
"The idea is to build this community globally, with the hope that as the model gets more generalized and more robust, it becomes a foundational machine-learning model for groundwater," he said.
