Click the link to read the article on the Water Education Foundation website (Matt Jenkins):

September 25, 2025

Western Water in-depth: After a thwarted quest to better predict the effects of drought and climate change, federal water managers are taking a radically different approach

After four years of contentious negotiations, the seven states that rely on water from the Colorado River are racing against the clock to reach agreement on a new long-term operating strategy for the river’s dams and reservoirs. They face a Nov. 11 deadline from U.S. Interior Department officials to signal whether they think a deal among them is likely.

This is a high-stakes moment on the Colorado: Some 40 million people, 5.5 million acres of farmland and a $1.4 trillion economy depend on water from the river. But the double whammy of climate change and a now-quarter-century-long drought has strained relationships between the seven states that share the dwindling river.

Over the past two decades, scientists, engineers and water managers have invested tremendous effort in trying to deduce what the future might bring. They have used reconstructions of climate patterns stretching more than 1,200 years into the past to understand natural variability, and turned to global models to better grasp the potential impacts of climate change.

A key player in the effort has been the federal Bureau of Reclamation, which is primarily responsible for operating the massive dam-and-reservoir system on the Colorado River. Its in-house research and computer modeling team has played a crucial role in bringing new science about climate variability and change to Colorado River water managers.

Even with that, though, water managers have been repeatedly blindsided after conditions on the river proved even worse than predicted. Two earlier rounds of negotiations, dating back to 2005, yielded a pair of “interim” operating agreements to help the states weather the drought. But the river’s flow has continued to deteriorate so rapidly that water managers have found themselves stuck in a perpetual scramble to buy themselves time before the river enters an all-out crisis.

“The policies weren’t robust enough, and we were in this Band-Aid mode,” says Carly Jerla, who heads Reclamation’s long-term planning process and was previously a leader on the research and modeling team. Everyone, she says, realized that “we need something else.”

As a result, Reclamation has quietly abandoned the effort to rely on best guesses about the river’s future via traditional modeling methods. Now, it’s bringing a radically different style of thinking to the negotiating table: Decision Making Under Deep Uncertainty, or DMDU.

The approach focuses on testing out operating strategies, with the help of artificial intelligence, that perform well against a far wider range of possible hydrologic scenarios than has ever been considered before — some of which no one on the river may anticipate or even be able to imagine. DMDU gives water managers a way to see how well their ideas fare, and to better understand how, and why, they might fail.

Scrambling to Stay Ahead of the Curve

Reclamation’s research and modeling team is based in Boulder, Colo., and works out of a nondescript University of Colorado building tucked between a city bus depot and an Audi dealership a mile from campus. The Reclamation team shares an office with the university’s Center for Advanced Decision Support for Water and Environmental Systems (CADSWES), which developed the software system used to model the Colorado.

Reclamation’s collaboration with CADSWES began in the mid-1990s, and was initially led by Terry Fulp, who would go on to serve as the agency’s regional director for the Lower Colorado River Basin. CADSWES provided modeling know-how, but it also served as a pipeline of talented grad students that its director, Professor Edie Zagona, would send Fulp’s way. Many of the most promising candidates wound up working for Fulp’s team, which operated with relative autonomy within Reclamation’s larger hierarchy.

“We kind of flew under the radar,” says Fulp, who retired in 2020. “We had a little bit of a notion that we were special. But we also didn’t want to be too special.”

As the team took shape, trouble was brewing on the river. The 1922 Colorado River Compact, which initially allocated the river’s water between the states, was based on an assumption that average annual flows on the river were 16.4 million acre-feet per year. Over the past century, however, that number has decreased by approximately 20 percent.

A dramatic wakeup call came in 2002, two years after the drought first took hold. Inflows to Lake Powell, one of the two main reservoirs on the river, were only about 25 percent of average, and water managers had the unnerving realization that the world might be changing in ways they couldn’t predict.

“We were walking into a complete unknown,” says Pat Mulroy, who at the time was the head of the Las Vegas-based Southern Nevada Water Authority. “You have to assume that a 2002 runoff is not an anomaly, but that it’s going to happen again, and it’s going to happen with greater frequency.”

In 2005, governors’ representatives from the seven states began to negotiate an operating strategy they hoped would give them a way to ride out the deepening drought. But they were treading into delicate territory.

Legal Minefields and Flawed Crystal Balls

The Colorado River is governed by a complex series of rules laid out not just by the Colorado River Compact, but by an amalgamation of subsequent laws, treaties, agreements and court decisions that are collectively known as the “law of the river.” That has set up fundamental tensions over how the river’s water is divided not just between individual states, but also — because of the Compact’s legal structure — between the Upper Basin states of Colorado, Utah, Wyoming and New Mexico and the Lower Basin states of California, Arizona and Nevada, as well as the U.S. and Mexico, which has its own share of the river’s water.

Numerous legal minefields lurk within the law of the river, ambiguous provisions about which various states deeply disagree. Among the thorniest are: What is the Upper Basin’s precise obligation to provide water to the Lower Basin downstream? What are the relative responsibilities of the Upper and Lower basins in ensuring that Mexico receives its legal entitlement to water? How does water that the Lower Basin uses from local tributaries factor into its Compact entitlement?

The negotiating effort that began in 2005 was an attempt to find creative ways to survive the drought while staying within the boundaries of the Compact. By avoiding those legal minefields, the states could capitalize on areas of mutual flexibility to meet everyone’s needs — or at least get as close as possible.

To figure out how to make it work, the states’ representatives and their technical support staff began relying on Reclamation’s research and modeling team in Boulder to calculate the probabilities of success or failure for various options they were considering. In 2007, the negotiating effort yielded a set of “interim guidelines” for Colorado River operations that would remain in effect until 2026.

During that process, Fulp and his colleagues had started using tree-ring based reconstructions of past climate history, together with computer projections of the possible impacts of climate change, to get a clearer sense of the future. But as the effort went on, the team’s members realized they had a problem: The results from the global climate models weren’t squaring with what they saw playing out in real time.

“The climate change projections in the Colorado didn’t map up with what we’ve been experiencing the last 10, 15, 20 years,” says Alan Butler, a research and modeling group chief on Reclamation’s Boulder team. “There was a disconnect.”

That disconnect only seemed to be getting worse. One set of climate projections, for instance, suggested that future flows on the Colorado could range from less than five million acre-feet a year to more than 45 million — twice as much water as came down the river in 1983 in a massive flood that nearly tore apart Glen Canyon Dam.

“That’s just a massive range,” says Nolie Templeton, a senior policy analyst for Central Arizona Project, which supplies water to cities like Phoenix and Tucson, as well as tribes. “If you get a five-million-acre-foot river, you’re going to be planning and adapting significantly differently than if the dam gets blown out because it’s 45.”

Jim Prairie, the other research and modeling group chief on Reclamation’s Boulder team, recalls a warning he got from a respected climate modeler in 2009: Global climate models are research, not decision-making tools. They were never intended to provide the kind of probability-based projections that water managers so desperately needed.

The team began to back off from its pursuit of long-term probabilities and search for a better approach.

Learning to Navigate Uncertainty 

Humans are practically hardwired to look to past experience to anticipate what the future might hold. Yet the world is changing in ways that our lived experience is ill-suited to help us comprehend. Decision Making Under Deep Uncertainty is a broad conceptual approach to addressing that problem.

Robert Lempert is a principal researcher at the RAND Corporation, the Santa Monica-based think tank that made its name devising Cold War nuclear deterrence strategy for the military. He’s also one of the intellectual pioneers of DMDU, a concept that’s being increasingly applied to long-term policy and planning challenges where future conditions are tough to predict. DMDU has been used in fields ranging from infrastructure, energy and transportation planning to public health and global security, and has helped cut airlines’ fuel costs and carbon emissions, formulate pandemic responses and analyze the effectiveness of the federal government’s terrorism risk insurance program.

It is particularly suited to situations where decision makers cannot reach consensus about future conditions or when traditional forecasting methods prove inadequate — exactly the problem that Reclamation’s team found itself facing with the climate models.

“What the climate models really give us,” Lempert says, “is overwhelming scientific evidence that the stable planning environment we built the system on has disintegrated.”

Rather than trying to make a best guess about what’s probable, DMDU is laser focused on what’s possible. A DMDU analysis typically starts by generating a wide range of possible future scenarios — or, in the case of a river, future flows. Policy makers can then test potential operating strategies to see which perform reasonably well, or are most robust, against that range. Based on those results, the operating strategies can then be refined to make them even stronger.

The process can also be used to identify vulnerabilities in the system and flag them with “signposts.” If system conditions begin approaching those danger zones, the people who depend on them can take up the challenge of devising contingency plans, or damage-control efforts, to stave off a descent into a full-blown water-supply crisis. Navigating those hazardous areas requires difficult choices, but flagging them up front — even if decision makers defer action on them to only when they absolutely have to be dealt with — allows for crucial wiggle room: They can still take some action in the face of uncertainty, even as they punt the really difficult questions to the future.

Lempert and other RAND researchers led much of DMDU’s conceptual development, and they occasionally crossed paths — and exchanged business cards — with members of Reclamation’s Boulder team. Then in 2009, when the team’s members began work on the Colorado River Basin Water Supply and Demand Study, a comprehensive look at the river’s next 50 years, they realized they needed help.

“We found ourselves buried in data,” says Jerla, who has headed the team since 2010. “And we were like, ‘Anyone got those RAND guys’ numbers to come dig us out of this mess?’”

A Brave New World

Even after the seven states reached agreement on the 2007 interim guidelines, the rapidly changing realities of the river forced them into a near-constant series of ongoing negotiations. In 2012, the Reclamation team brought RAND representatives to the meetings to familiarize the states’ technical staff with DMDU.

That effort — at least initially — wasn’t exactly a smashing success. The states’ water managers were flummoxed by RAND researchers expounding on abstract concepts from the world of decision science. And, Jerla says with a laugh, “I don’t know that any of usreally even understood what was happening.”

The partnership between Reclamation and RAND wound down after the Water Supply and Demand Study concluded. But the Reclamation team continued working to incorporate DMDU techniques into its research and modeling.

At Reclamation’s behest, Zagona, University of Colorado professor Joseph Kasprzyk and others on the CADSWES team took the Colorado River model and married it with an AI tool called a “multi-objective evolutionary algorithm” developed at Penn State. The algorithm — somewhat ominously named Borg — is a sort of computational supercharger that can create many potential operating strategies, test them out in the river model, and sort through them to find the ones that perform best.

In 2016, the Reclamation team began exploratory work with the Borg-enhanced software to see what it could do. The following year, Kasprzyk, Zagona and a graduate student named Elliot Alexander — who would quickly be hired on with the Reclamation team — used the augmented modeling package to find an operating strategy for Lake Mead, the other main reservoir on the Colorado, that outperformed the one the states had painstakingly negotiated for the 2007 interim guidelines.

But the operation of Lake Mead is just one, albeit very important, variable in the complex Colorado River system. The potential beauty of Borg was that it can combine many policy variables to identify strategies that perform well across multiple objectives in a wide range of hydrologic scenarios.

There’s a catch, however: Multi-objective strategies, practically by definition, demand constant compromise. Keeping the water level in Lake Powell as high as possible, for example, improves the odds of being able to continue generating hydropower at Glen Canyon Dam. But it simultaneously limits water deliveries to the downstream states of California, Arizona and Nevada, among other tradeoffs.

Still, Borg offered a little more. The “evolutionary” part of the algorithm gave it the ability to essentially breed well-performing operating strategies with each other — and even artificially induce mutations — to create new approaches that might perform even better.  

Yet Borg sometimes showed a naughty streak.

“It would find a lot of mathematical solutions that maybe were optimal for a certain metric,” says Butler. “But then you’d look at them and you’d think: ‘That’s just absurd.’”

In one test, the team set Borg loose on a mission to minimize the frequency of water shortages over a 30-year model run. The algorithm diligently avoided implementing water-delivery cuts for as many years as possible, until Lake Mead dropped so low that water could not be released from the reservoir, resulting in a sudden, six-million-acre-foot cut to California, Arizona and Nevada — an amount roughly equal to those three states’ entire annual Colorado River water use.

Ultimately, both Reclamation and the state and local water managers would end up using Borg not to generate specific strategies for consideration, but to test strategies of their own devising. But the exploratory work with Borg helped create a virtual anvil on which they could hammer out their own strategies and see how they compared with the bigger world of possibilities — even though some of those might be absurd.

“Borg created this dartboard where, if we’re throwing darts, at least we know where they land,” says Rebecca Smith, Reclamation’s Lower Colorado Basin research and modeling team lead. “Without having that, we’re just saying: ‘I guess this is good’ — but we don’t know how much better we could do.”

Translating Science into Action

Meanwhile, the clock was ticking on the Colorado River. After six grueling years of negotiations, the states reached agreement in 2019 on a Drought Contingency Plan that added to the interim guidelines. But the entire package of agreements was set to expire in just another six years. And so, in 2021, the state negotiating teams started meeting informally again to develop what, after a decade and a half of workarounds, they hoped would be a longer-term operating strategy.

While that was happening, the Reclamation team tasked Nathan Bonham, a newly arrived University of Colorado doctoral student who would also eventually be hired by Reclamation, with refining the methods used to assess system vulnerabilities and the robustness of potential operating strategies. That work led to a public web tool, designed in collaboration with CADSWES and consulting firm Virga Labs, that would put the DMDU-inspired upgraded software package into the hands of the negotiating teams as well as water agencies and anyone else, like tribes and environmental groups, with an interest in the river’s future.

The effort to develop the web tool reached a blistering pace over six months in 2023. Smith and H.B. Zeff, another Reclamation engineer at the time, would upload massive numbers of simulations to Microsoft’s cloud of high-performance Azure computers and remotely babysit the models as they ran, only to discover that the computers were rebooting themselves to install updates in the middle of the night.   

Despite such glitches, the upgraded software package went online in November 2023, just as the negotiating effort to develop a post-2026 operating strategy was kicking into high gear. Now, water users had a way to test the strategies they were considering against 8,400 possible hydrologic scenarios.

One of the biggest challenges is presenting such complex data in a way that allows negotiators to compare the tradeoffs between various operating strategies.

“I can crunch the numbers all day long,” says Bonham, “but there’s a whole other element of how do you present it visually?”

In the web tool, each strategy under consideration can be displayed on an interactive parallel-axis chart. To a first-time user, the charts look like twisted skeins of yarn on a loom gone haywire. But with familiarity over time, they become a window into possibility.

Users of the web tool can adjust the relative importance of various “performance objectives”: water levels at lakes Mead and Powell; water releases from the Upper Basin downstream to the Lower Basin; potential water cuts to Lower Basin states; favorable conditions for native fish in the Grand Canyon. Then, at least theoretically, they can find strategies that help them meet the goals they most care about without adversely affecting the objectives of other users, whose buy-in they need for a real-world agreement.

The web tool’s vulnerability analyses also help identify the danger zones — like low river flows below which problems start to occur at particular points in the system — that would necessitate more extensive damage-control efforts.

“That puts some numerical context around it,” Prairie says, “to track not just a feeling, but actually a level of flow that the analysis shows is a point where you start to see failure.”

DMDU’s ability to accurately flag those hazards could also potentially help water managers better respond when conditions start getting really bad.

“If we can understand where (an operating strategy) falls short, and have also seen what is more effective if things get worse,” says Smith, “then we are more prepared to adapt.”

Crunch Time for a Deal

The governors’ representatives are now racing to meet the Nov. 11 deadline to notify the Interior Department whether they’re likely to reach agreement on a post-2026 operating strategy. Reclamation’s Boulder team has been busy helping them with on-the-spot modeling work.

For water managers, DMDU is proving to be a mixed blessing — or a double-edged sword. It is helping illuminate and more quantitively delineate the hazardous areas in the river’s future. But it’s also pushing hard questions to the fore.

“It’s a totally different way to think about risk,” says Central Arizona’s Project’s Templeton. “Just by exploring all these potentials, we’re understanding that there are critical thresholds in our future that should prompt some decision-making. That definitely has resonated within our agency.”

The catch, she says, is that DMDU doesn’t provide an unequivocal path through those decisions; it only illuminates the tradeoffs.

“The DMDU approach doesn’t say ‘yes’ or ‘no’ to any of those,” she says. “It’s always: ‘It depends.’”

The algorithm is not going to find a super-strategy for the future — at least not one that all seven states can agree to.

“I think many people like the idea of being able to have a magic strategy. But on the ground, it’s not that simple,” says Laura Lamdin, a senior engineer with the Metropolitan Water District, which supplies urban Southern California. “Having the ability to quickly test a bunch of ideas as you try and incorporate some out-of-the-box thinking is valuable to creating those more handcrafted strategies.”

In the end, DMDU’s real utility may not lie in delivering miracle fixes, but simply in helping water managers better understand the ramifications of their decisions.

The negotiators for the states may be able to reach agreement on a less-than-perfect plan that still gives them the flexibility to deal with tougher questions as they arise. In fact, it seems likely that any operating strategy the states can agree on will follow the incremental approach they’ve taken so far. If that turns out to be true, DMDU could help bring a better-informed style of incrementalism to the effort to work through the problems on the river.

In that mode of problem-solving, the danger zones are critical. In one sense, they are the perilous realms where water gets really tight. Yet they also mark the legal minefields that the states have so carefully steered clear of throughout the negotiations since 2005.

“One of the big problems is there’s a lot of the Compact questions that have been put off for many, many, many years,” says J.B. Hamby, the California governor’s representative in the negotiations. “We’ve continued to dance around them — and (now) here we are dealing with them, but with really bad hydrology, which then puts these core questions to the test.”

Paradoxically, as punishing as the entire two-decade-long negotiating process has been, it has spurred an era of innovation on the river, opening the door to more flexible reservoir operations and what has grown to be a massive water banking and transfer program.  

Viewed more optimistically, then, DMDU’s ability to mark the danger zones in a post-2026 operating strategy might also reveal places where there could be new opportunities for the states to cut even more of the incremental deals they’ve managed to make between themselves so far.

Tough Choices Lie Ahead

Still, nearly everyone at the negotiating table acknowledges that a hard reality lies behind all of this. Annual water use throughout the Colorado River Basin currently exceeds inflows by at least 3.6 million acre-feet. The only way to make the numbers work over the long term — to truly make the Colorado River system robust against a future in which the only certainty is that there will be far less water — is to reduce the total amount of water used throughout the entire basin.

Depending on how big they are, water cuts could have enormous economic impacts. In fact, the biggest point of contention in the negotiation of the post-2026 operating guidelines is which states would take cuts, and how big they’d be. In 2024, California, Arizona and Nevada committed to collectively reducing their use by 1.25 million acre-feet a year — 20 percent of what they used that year — and proposed splitting additional cuts with the Upper Basin and Mexico up to a total of 3.9 million acre-feet.

For their part, Colorado, Utah, Wyoming and New Mexico have, at least publicly, been adamant about not taking any cuts. They argue that, without any large upstream reservoirs backstopping their water supplies, they’ve already been disproportionately affected by drought and climate change — and, because they’ve grown slower than their downstream counterparts, they’re still entitled to water under the Compact that they haven’t yet put to use. 

Breaking through that stalemate is the key challenge negotiators now face, and by most accounts their prospects for doing so are dim. But regardless of whether they can resolve that impasse by November, the really hard questions may be coming sooner rather than later.

The research and modeling team’s analyses suggest that when the Colorado River’s 10-year average annual flow dips into the 12- to 13-million acre-foot range, a lot of things start going wrong. As it happens, the river’s flows over the past five years have fallen squarely within that range. And in September, an independent group of Colorado River experts released an analysisshowing that, without immediate reductions in water use, the amount of “realistically accessible storage” in Lake Powell and Lake Mead could essentially be exhausted by early 2027.     

The 21^st century Colorado River is a world of inescapable tradeoffs, and DMDU is, at root, a search for the least-bad strategy to which everyone can agree. But, Smith says, that kind of compromise comes with a big question: “Are we prepared to deal with the realities of whatever gets chosen?”

“That’s the thing about DMDU,” she adds. “It shifts when you have to make the call — but you do still have to make a call.”

Reach Writer Matt Jenkins at mjenkins@watereducation.org. 

Know someone who wants to stay connected to water in the West? Encourage them to sign up for Western Water and follow us on LinkedIn, X (formerly Twitter), Facebook and Instagram.