The wetland puzzle that stumped hydrology for decades—how physics and AI joined forces to predict unmeasured regions

https://phys.org/news/2026-03-wetland-puzzle-stumped-hydrology-decades.html

Home
Nanotechnology
Physics
Earth
Astronomy & Space
Chemistry
Biology
Other Sciences
10 hours ago
Earth Earth Sciences
Earth Environment
The wetland puzzle that stumped hydrology for decades—how physics and AI joined forces to predict unmeasured regions
by Ali Ameli

edited by Sadie Harley, reviewed by Robert Egan

Editors’ notes
dialog
The GIST
Add as preferred source
The wetland puzzle that stumped hydrology for decades — how physics and AI joined forces to predict unmeasured regions
Conceptual overview of pothole-controlled runoff generation and how our hybrid framework represents it. The top row illustrates the threshold sequence in depression-dominated landscapes: water first fills isolated potholes, then spills once local storage is exceeded, and finally forms connected flow paths that rapidly route water to streams. The bottom row shows how a physics-informed learning step estimates key, hard-to-measure parameters controlling these thresholds (e.g., effective storage capacity), enabling the hydrologic model to reproduce both storage dynamics and the resulting streamflow behavior across space. Credit: Ali Ameli

For years, the Prairie Pothole Region has bothered me in a very specific way. On a map, it looks like a normal landscape: fields, gentle slopes, small streams. But hydrologically, it behaves like something else entirely. The surface is peppered with countless depressions—wetlands and “potholes”—that can store water for days, months, or even years. Most of the time, rainfall and snowmelt do not move cleanly downhill into channels. They disappear into storage. Then, sometimes, they don’t.

A wetland fills. It spills into a neighboring depression. A chain of depressions connects. And suddenly, a watershed that looked “disconnected” on Monday behaves like a fast-draining basin on Tuesday. Streamflow can jump from near-zero to a flood pulse with very little warning. The same watershed can act like two different systems depending on how full the potholes are.

That threshold-driven behavior—fill, then spill, then connect—is why this region is among the most difficult places on Earth to predict streamflow. Traditional conceptual models struggle because they were not built around thousands of small, interacting storage elements.

Purely data-driven models can fit historical patterns, but they often fail when asked to transfer to new basins or new periods, especially in a region where the governing “state” is storage, and storage is rarely observed well.

This left us with a question that sounds simple but turns out to be hard: can we build a model that predicts streamflow in this landscape for the right reasons and can be transferred to watersheds where we do not have long streamflow records?

Our new paper, published in the journal Water Resources Research, is our attempt to answer that.

Why this matters
People often assume hydrologic prediction is mainly about precipitation. If we know how much rain falls, we can estimate how much water will show up in the river. In the Prairie Pothole Region, that assumption breaks down.

Here, what matters is not just input (rain/snow) but the internal configuration of storage. Two years with similar precipitation can produce completely different runoff depending on how full the pothole network already is. That is exactly the kind of nonlinearity that creates surprises—sudden flooding, unexpected lack of flow, and big interannual swings that are hard to anticipate.

The stakes have grown. The Prairie Pothole Region is North America’s agricultural heartland—the breadbasket that helps feed the continent. Over half of its wetlands have been drained for farming, removing natural sponges that once absorbed floodwaters.

The wetland puzzle that stumped hydrology for decades—how physics and AI joined forces to predict unmeasured regions
Geographic location of the Prairie Pothole Region (PPR), spanning parts of Canada (Alberta, Saskatchewan, Manitoba) and the United States (Montana, North Dakota, South Dakota, Minnesota, Iowa). The red boundary delineates the region’s extent across both countries. The blue pixels represent long-term inundated areas (potholes and lakes). Credit: Water Resources Research (2026). DOI: 10.1029/2025wr040280

As climate change intensifies precipitation extremes, communities across the northern Great Plains face growing flood risk with fewer tools to anticipate it. And the vast majority of watersheds in the region have no stream gauges. Prediction in these unmeasured landscapes—where managers actually need forecasts—has remained essentially unsolved for decades.

If we cannot predict when watersheds will connect, we struggle to forecast floods. If we cannot represent how water is retained and later released, we struggle to anticipate nutrient pulses and water-quality impacts. And if we cannot generalize across space, we cannot provide useful information in the many places where observations are sparse or absent.