larmedias - 科学技術メディア

AI Models Map the Colorado River’s Hard Choices

Apr 8, 2026, 2:00 PM

出典: IEEE Spectrum AI

<img src="https://spectrum.ieee.org/media-library/overhead-view-of-horseshoe-bend-an-incised-meander-shaped-like-the-letter-u.jpg?id=65487612&width=1245&height=700&coordinates=0%2C156%2C0%2C157"/> The <a href="https://spectrum.ieee.org/tag/colorado-river" target="_blank">Colorado River</a> begins as snow. Every spring, the mountain snowpack of the Rockies melts into streams that feed into reservoirs that supply 40 million people across seven U.S. states. The system has worked, more or less, for a century. That century is over.By some measures, 2026 is shaping up to be the worst year the river has seen since records began. Flows are down <a href="https://www.science.org/doi/10.1126/science.abj5498" rel="noopener noreferrer" target="_blank">20 percent from 2000 levels</a>. Lake Powell, the reservoir straddling Utah and Arizona, may drop below the threshold for generating hydropower <a href="https://coloradosun.com/2026/02/18/lake-powell-forecast-critical-lows-federal-study/" rel="noopener noreferrer" target="_blank">before the year is out</a>. The negotiations between the seven states over how to <a href="https://www.nytimes.com/2026/02/13/climate/colorado-river-cooperation-missed-deadline.html" rel="noopener noreferrer" target="_blank">share what’s left have collapsed twice</a>, and the U.S. federal government is threatening to impose its own plan.While the states argue and the river shrinks, a growing set of machine learning tools is being deployed across the basin. Federal water managers are running millions of simulations to stress-test reservoir strategies against different possible futures. Researchers are forecasting streamflow months out using satellite data and deep learning. These technologies don’t promise to resolve the crisis, but they’re making the tradeoffs visible. They’re showing, more precisely than ever before, what each decision will cost.<h2>Seeing Further into the River’s Future</h2>Nobody manages more of the Colorado River’s daily operations than the <a href="https://www.usbr.gov/" rel="noopener noreferrer" target="_blank">U.S. Bureau of Reclamation</a>. If the federal government follows through on its threat to impose a water-sharing plan, it will be Reclamation doing the imposing, and making decisions about how much water flows from Lake Powell and Lake Mead, the two largest reservoirs in the country. The agency is not new to sophisticated modeling. For years, Reclamation’s researchers have combined paleoclimate reconstructions, global circulation models, and scenario planning to predict the river’s future. Machine learning tools are extending that toolkit, says <a href="https://www.linkedin.com/in/chris-frans-phd-pe-74491579/" target="_blank">Chris Frans</a>, Reclamation’s water availability research coordinator, and are already informing real operational decisions.The clearest gains are in streamflow forecasting. Machine learning techniques, pulling on satellite data and weather stations well outside the basin, now outperform traditional methods across a range of conditions. Forecasts update every hour. In some areas, managers are getting five to seven days of advance warning on flood events, compared with three in the past, which gives them time to reduce the water in reservoirs before high inflows arrive.The scale of scenario modeling has also expanded dramatically. A decade ago, running 100,000 individual simulations was a landmark study. Now, says <a href="https://www.colorado.edu/cadswes/alan-butler" rel="noopener noreferrer" target="_blank">Alan Butler</a>, who manages Reclamation’s research and modeling group for the lower Colorado Basin, millions of simulations feed the analytical tools used in the current guidelines. Those simulations map out how different operating strategies perform across widely varying futures—making the tradeoffs between them harder to ignore.<h2>Dividing a Shrinking River</h2>Knowing how much water is coming is one problem. Deciding who gets it is another. At the center of that process is the <a href="https://coloradoriverscience.org/Colorado_River_Simulation_System_(CRSS)" rel="noopener noreferrer" target="_blank">Colorado River Simulation System</a> (CRSS), which models how water moves through the basin’s reservoirs, canals and pipelines under more than a century of legal and regulatory constraints. This Reclamation model is an imperfect representation, but it has been the foundation of river negotiations for decades.A tool called <a href="https://riverware.org/" rel="noopener noreferrer" target="_blank">RiverWare</a>, first developed in the early 1990s at the University of Colorado Boulder, lets states, cities, and tribes run their own scenarios through CRSS. Before RiverWare, these groups didn’t have confidence in Reclamation’s numbers, says <a href="https://www.colorado.edu/ceae/edith-zagona" rel="noopener noreferrer" target="_blank">Edith Zagona</a>, a Boulder professor who directs the <a href="https://www.colorado.edu/cadswes/" rel="noopener noreferrer" target="_blank">Center for Advanced Decision Support for Water and Environmental Systems</a>, the center that built it. “There was just this huge lack of trust.” The solution was letting stakeholders inspect the assumptions built into the Riverware model—how much water was available, how it could be used, and under what rules. Getting stakeholders to trust the model turned out to be the easier problem. The harder one is what to do when the model itself can’t predict a single probable future. That question drove Zagona toward a framework called decision making under deep uncertainty, which trades prediction for stress-testing policies against thousands of possible futures.The tool Zagona’s group developed with Reclamation and the consulting firm <a href="https://virgalabs.io/" rel="noopener noreferrer" target="_blank">Virga Labs</a> puts the framework into practice in a web-based tool, running CRSS across more than 8,000 possible future water supply scenarios to show how different management strategies hold up against the full range of what climate change might bring. At its center is an evolutionary algorithm called Borg, which generates and iteratively refines those strategies, searching for plans that perform well across many scenarios. The result is a set of tradeoffs, not a single answer. <a href="https://riverware.org/riverware/ugm/2025/PDFs/Users/10.kasprzyk-rwugm2025-borgRW-full.pdf" rel="noopener noreferrer" target="_blank">Borg-RiverWare</a> has already shaped the ongoing negotiations over the river’s next operating rules, generating the scenarios and data that Reclamation used in its modeling tools. Those tools give stakeholders a common analytical foundation for negotiations. Now Zagona’s center is pushing the approach further. A system in development would let negotiating parties test competing proposals on the fly, showing how one side’s policy choices would ripple through the system and identifying areas of potential compromise during the negotiation itself.<h2>New Tools for Forecasting the Colorado</h2>Reclamation and Zagona’s center aren’t the only ones trying to see further into the river’s future. At Metropolitan State University of Denver, a team led by <a href="https://red.msudenver.edu/expert/mohammad-valipour/" rel="noopener noreferrer" target="_blank">Mohammad Valipour</a> has been building a forecasting system that uses deep learning to issue drought warnings across seven rivers in Colorado, from seven days to six months out. In a region where ground gauges are sparse and mountains make installation difficult, the team found that NASA satellite data outperformed in-field measurements. The goal, Valipour says, is a statewide drought alarm system that gives farmers and water managers more time to respond.At Utah State University, <a href="https://engineering.usu.edu/cs/directory/faculty/boubrahimi-filali-soukaina" rel="noopener noreferrer" target="_blank">Soukaina Filali Boubrahimi</a> is attacking a different problem: how conditions at one point in the river ripple downstream weeks later. Using a graph neural network that treats each monitoring station as a node, her team built a map of the river’s interdependencies across one of the most contested water systems in the world. She says the approach could extend to other overtaxed basins.“If you can figure out the Colorado River,” she says, “anyone else dealing with a stressed river system is going to be interested in what you learned.” <img alt="Line graph of snow water equivalent projection in the upper Colorado region. As of March 26th 2026, most maximum projections through June still fall short of the median." class="rm-shortcode" data-rm-shortcode-id="51c77aad5a41503f03b2d3ade20069a1" data-rm-shortcode-name="rebelmouse-image" id="07b94" loading="lazy" src="https://spectrum.ieee.org/media-library/line-graph-of-snow-water-equivalent-projection-in-the-upper-colorado-region-as-of-march-26th-2026-most-maximum-projections-thr.jpg?id=65487654&width=980"/> Snowpack in the upper Colorado River basin is far below normal. As of late March 2026, measurements across 130 sites were about 35 percent of the median, with projections showing continued shortfalls.<a href="https://nwcc-apps.sc.egov.usda.gov/awdb/basin-plots/Proj/WTEQ/assocHUC2/14_Upper_Colorado_Region.html?hucFilter=14" target="_blank">USDA Natural Resources Conservation Service (NRCS)</a><h2>What the Models Can’t See</h2>Across the basin, researchers and water managers are running into the same wall. The models learn from historical data, but that data describes a river that no longer exists. Valipour found that feeding his models only the last decade outperformed using longer records. Filali Boubrahimi’s model struggles most in drought conditions, precisely when predictions matter most, because recent prolonged droughts don’t resemble the historical training data. One workaround is to train models on data from basins that have already experienced what the Colorado hasn’t yet.Even so, better forecasts do not resolve the central problem. While the tools can show you what a drier future looks like across a thousand possible scenarios, they can’t tell you who should bear the cost of it. The cuts coming to the basin are going to be enormous, says <a href="https://www.colorado.edu/center/gwc/brad-udall" rel="noopener noreferrer" target="_blank">Brad Udall</a>, a water and climate research scientist at Colorado State University’s <a href="https://watercenter.colostate.edu/" target="_blank">Colorado Water Center</a>, and will fall mostly on <a href="https://spectrum.ieee.org/tag/agriculture" target="_blank">agriculture</a>. They may fundamentally reshape communities that have built their economies around water for generations. “AI has no business being in the realm of replacing human values and human judgments,” he says.The tools, by most measures, are doing exactly what they were built to do: The negotiating parties understand what is coming, and they are not disputing the projections. Zagona, who has worked on the Colorado River for 45 years, sees reasons for optimism. “The tools are bringing people to the table,” she says. “They’re at the table arguing. But at least they’re at the table.”

Details

The Colorado River begins as snow. Every spring, the mountain snowpack of the Rockies melts into streams that feed into reservoirs that supply 40 million people across seven U.S. states. The system has worked, more or less, for a century. That century is over.

By some measures, 2026 is shaping up to be the worst year the river has seen since records began. Flows are down 20 percent from 2000 levels. Lake Powell, the reservoir straddling Utah and Arizona, may drop below the threshold for generating hydropower before the year is out. The negotiations between the seven states over how to share what’s left have collapsed twice, and the U.S. federal government is threatening to impose its own plan.

While the states argue and the river shrinks, a growing set of machine learning tools is being deployed across the basin. Federal water managers are running millions of simulations to stress-test reservoir strategies against different possible futures. Researchers are forecasting streamflow months out using satellite data and deep learning. These technologies don’t promise to resolve the crisis, but they’re making the tradeoffs visible. They’re showing, more precisely than ever before, what each decision will cost.

Seeing Further into the River’s Future

Nobody manages more of the Colorado River’s daily operations than the U.S. Bureau of Reclamation. If the federal government follows through on its threat to impose a water-sharing plan, it will be Reclamation doing the imposing, and making decisions about how much water flows from Lake Powell and Lake Mead, the two largest reservoirs in the country.

The agency is not new to sophisticated modeling. For years, Reclamation’s researchers have combined paleoclimate reconstructions, global circulation models, and scenario planning to predict the river’s future. Machine learning tools are extending that toolkit, says Chris Frans, Reclamation’s water availability research coordinator, and are already informing real operational decisions.

The clearest gains are in streamflow forecasting. Machine learning techniques, pulling on satellite data and weather stations well outside the basin, now outperform traditional methods across a range of conditions. Forecasts update every hour. In some areas, managers are getting five to seven days of advance warning on flood events, compared with three in the past, which gives them time to reduce the water in reservoirs before high inflows arrive.

The scale of scenario modeling has also expanded dramatically. A decade ago, running 100,000 individual simulations was a landmark study. Now, says Alan Butler, who manages Reclamation’s research and modeling group for the lower Colorado Basin, millions of simulations feed the analytical tools used in the current guidelines. Those simulations map out how different operating strategies perform across widely varying futures—making the tradeoffs between them harder to ignore.

Dividing a Shrinking River

Knowing how much water is coming is one problem. Deciding who gets it is another. At the center of that process is the Colorado River Simulation System (CRSS), which models how water moves through the basin’s reservoirs, canals and pipelines under more than a century of legal and regulatory constraints. This Reclamation model is an imperfect representation, but it has been the foundation of river negotiations for decades.

A tool called RiverWare, first developed in the early 1990s at the University of Colorado Boulder, lets states, cities, and tribes run their own scenarios through CRSS. Before RiverWare, these groups didn’t have confidence in Reclamation’s numbers, says Edith Zagona, a Boulder professor who directs the Center for Advanced Decision Support for Water and Environmental Systems, the center that built it. “There was just this huge lack of trust.” The solution was letting stakeholders inspect the assumptions built into the Riverware model—how much water was available, how it could be used, and under what rules.

Getting stakeholders to trust the model turned out to be the easier problem. The harder one is what to do when the model itself can’t predict a single probable future. That question drove Zagona toward a framework called decision making under deep uncertainty, which trades prediction for stress-testing policies against thousands of possible futures.

The tool Zagona’s group developed with Reclamation and the consulting firm Virga Labs puts the framework into practice in a web-based tool, running CRSS across more than 8,000 possible future water supply scenarios to show how different management strategies hold up against the full range of what climate change might bring. At its center is an evolutionary algorithm called Borg, which generates and iteratively refines those strategies, searching for plans that perform well across many scenarios. The result is a set of tradeoffs, not a single answer.

Borg-RiverWare has already shaped the ongoing negotiations over the river’s next operating rules, generating the scenarios and data that Reclamation used in its modeling tools. Those tools give stakeholders a common analytical foundation for negotiations. Now Zagona’s center is pushing the approach further. A system in development would let negotiating parties test competing proposals on the fly, showing how one side’s policy choices would ripple through the system and identifying areas of potential compromise during the negotiation itself.

New Tools for Forecasting the Colorado

Reclamation and Zagona’s center aren’t the only ones trying to see further into the river’s future. At Metropolitan State University of Denver, a team led by Mohammad Valipour has been building a forecasting system that uses deep learning to issue drought warnings across seven rivers in Colorado, from seven days to six months out. In a region where ground gauges are sparse and mountains make installation difficult, the team found that NASA satellite data outperformed in-field measurements. The goal, Valipour says, is a statewide drought alarm system that gives farmers and water managers more time to respond.

At Utah State University, Soukaina Filali Boubrahimi is attacking a different problem: how conditions at one point in the river ripple downstream weeks later. Using a graph neural network that treats each monitoring station as a node, her team built a map of the river’s interdependencies across one of the most contested water systems in the world. She says the approach could extend to other overtaxed basins.

“If you can figure out the Colorado River,” she says, “anyone else dealing with a stressed river system is going to be interested in what you learned.”

Line graph of snow water equivalent projection in the upper Colorado region. As of March 26th 2026, most maximum projections through June still fall short of the median. Snowpack in the upper Colorado River basin is far below normal. As of late March 2026, measurements across 130 sites were about 35 percent of the median, with projections showing continued shortfalls.USDA Natural Resources Conservation Service (NRCS)

What the Models Can’t See

Across the basin, researchers and water managers are running into the same wall. The models learn from historical data, but that data describes a river that no longer exists. Valipour found that feeding his models only the last decade outperformed using longer records. Filali Boubrahimi’s model struggles most in drought conditions, precisely when predictions matter most, because recent prolonged droughts don’t resemble the historical training data. One workaround is to train models on data from basins that have already experienced what the Colorado hasn’t yet.

Even so, better forecasts do not resolve the central problem. While the tools can show you what a drier future looks like across a thousand possible scenarios, they can’t tell you who should bear the cost of it. The cuts coming to the basin are going to be enormous, says Brad Udall, a water and climate research scientist at Colorado State University’s Colorado Water Center, and will fall mostly on agriculture. They may fundamentally reshape communities that have built their economies around water for generations. “AI has no business being in the realm of replacing human values and human judgments,” he says.

The tools, by most measures, are doing exactly what they were built to do: The negotiating parties understand what is coming, and they are not disputing the projections. Zagona, who has worked on the Colorado River for 45 years, sees reasons for optimism. “The tools are bringing people to the table,” she says. “They’re at the table arguing. But at least they’re at the table.”

AI Models Map the Colorado River’s Hard Choices

Details

Seeing Further into the River’s Future

Dividing a Shrinking River

New Tools for Forecasting the Colorado

What the Models Can’t See

Related Knowledge

Machine Learning in Environmental Science

Water Resource Management