The agency's machine learning model, built from three decades of atmospheric sensor data, correctly flagged 93% of flash flood events in retrospective tests.
The phone alert that warns a hiker away from a flooded wash, the Wireless Emergency Alert that pulls a driver off a low-water crossing, the National Weather Service forecaster who flips a watch box from yellow to red: each one is the end of a chain that starts with a sensor and ends with a person who has to believe the warning. NASA's newest machine-learning system, built to catch flash floods earlier than current tools, has just shown it can hold up its end of that chain in simulation. The harder problem is the middle.
The system, called TACLS — Tracking and Analyzing Convergence in Large-scale Signals — produced forecasts in as little as 15 minutes and, in retrospective tests on severe weather from 2017 through 2023, captured 93% of the flash flood warnings that were actually issued (NASA Science). Those numbers come from a model trained on more than 30 years of ground-based sensor data that measures how much water vapor is sitting in the lower atmosphere. When the model sees the signal jump, it flags the moment as a candidate flash-flood precursor, sorts flagged events from data noise, and passes only the real-looking ones to a visualization tool reviewed by a human forecaster, who decides whether to act.
The 93% figure is a lab result. It was scored against historical events, not against a live warning feed. The 15-minute lead time is real but conditional: it depends on a meteorologist being at the screen, on a dashboard or chat reaching them in time, and on a downstream alert system that can push to a phone. The NASA team is clear that the design is human-in-the-loop, with the model flagging and prioritizing while meteorologists retain final warning authority.
What changes about the warning someone actually receives, then, depends less on the model and more on the pipeline. National Weather Service forecasters in Southern California are working to fold TACLS into their existing flash-flood forecasting workflow, according to the NASA release, but the project is in integration rather than production deployment. The principal investigator, Yehuda Bock, is a Distinguished Researcher at UCSD's Scripps Institution of Oceanography, and the work is a partnership between NASA, the agency's Jet Propulsion Laboratory, NOAA, and the NWS, funded through NASA's Earth Science Technology Office.
This is also where the equity question lives. Flash-flood fatalities are not distributed evenly across the map. They cluster in basins that are ungauged or poorly gauged, in lower-income neighborhoods built near concrete channels, in canyons used by outdoor workers, and at night, when people are asleep and the only available warning is a phone they may not be carrying. A model that buys 15 minutes is only as useful as the alert system that delivers it, and that delivery has long been the thinnest link in the chain.
The model itself is a public project. TACLS reuses several JPL-built technologies, including a Mars-mission visualization tool repurposed for flood forecasting, and the software and training data are slated for open-source release, which would let other agencies and researchers retrain or fork the system. The project is documented on NASA TechPort as project 117321.
What to watch: whether NWS integration moves from pilot to operational use in additional forecast offices, whether the model gets tested in the kinds of basins where current warnings are weakest, and whether the open-source release draws outside researchers into the loop. The next dominoes are forecaster training, alert-pipeline integration, and the harder, slower work of getting a verified warning to the person standing in the wash.