The Prairie Pothole Region — stretching from the Dakotas through the Canadian prairies — has defeated hydrological models for decades. The landscape contains thousands of small wetland depressions, each a few meters to a few hectares, scattered across glacially carved terrain. Classical hydrology treats these as noise: small ponds that fill and overflow, feeding into larger streams in a predictable cascade.
The reality is threshold-driven. Individual potholes fill to a critical level, then suddenly connect to their neighbors through surface overflow. Below threshold, each basin is isolated — water goes in, stays in, evaporates. Above threshold, basins merge into connected networks that drain as if the landscape had reorganized itself. The switch between disconnected and connected states is sharp, not gradual. The watershed blinks between two configurations.
Ameli's team at UBC embedded wetland storage physics into a deep learning model. The key was making the neural network respect the fill-and-spill dynamics: water accumulates in discrete depressions, each with a specific storage capacity, and connectivity emerges only when storage exceeds capacity. Training the model with this physical constraint doubled its transferability to ungauged basins — watersheds where no historical flow data exists.
The prairie pothole landscape is a distributed memory system. Each depression stores water independently. The stored water encodes information about recent rainfall history. The collective state of all basins — how full, how close to spilling, how recently connected — determines whether the next rainfall event produces runoff or disappears. The landscape doesn't respond to rain. It responds to rain in the context of everything it's already holding.
The “noise” of thousands of small ponds was the signal. The behavior that defied simple precipitation-to-runoff models was the model.