The GPS Signal That Could Make AI Forecasts Better at Predicting Catastrophic Rain

The same satellite signals that help you navigate with your phone might soon help predict devastating floods before they happen. That's the exciting possibility from new research published by scientists including Leonardo Trentini and his team, who found that ordinary GPS signals contain hidden information about extreme rainfall.

Here's how it works: when water vapor hangs heavy in the atmosphere, it actually slows down signals from Global Navigation Satellite Systems—your basic GPS satellites—by a tiny, measurable amount. Scientists call this delay the Zenith Wet Delay, or ZWD. Weather experts have known about this phenomenon for decades, but it turns out nobody had tried feeding this information into the powerful AI models that now forecast our weather.

Trentini and colleagues were the first to plug ZWD data into Aurora, one of the most advanced AI weather models currently being developed. The results were striking. When the model learned to read these GPS signal delays, it became significantly better at predicting heavy rain—the kind that causes dangerous flash floods and widespread damage.

At the most extreme rainfall levels—the kind meteorologists call the 99th percentile, meaning the worst 1% of storms—the model improved by 8.8% in what's called the Equitable Threat Score. That's a key measurement of how well forecasters can predict where dangerous weather will actually hit. Beyond raw scores, the researchers noticed something even more encouraging: the model's simulation of large-scale weather patterns became more realistic, matching what actually happens in nature more closely.

This matters because right now, even the smartest AI weather systems tend to underestimate severe precipitation. They miss the intensity of approaching storms or misjudge exactly where the worst rain will fall. That can mean the difference between a town getting hours of warning or just minutes.

The study suggests that GNSS observations—signals already being collected across thousands of weather stations worldwide—encode valuable information that machine learning models can actually learn to use. Best of all, this technology already exists. Weather agencies aren't waiting for new satellites; the data is already flowing. What Trentini's team showed is that scientists simply haven't been using it properly.

The research was submitted to Geophysical Research Letters and represents a collaboration between atmospheric physics, machine learning, and geophysics experts. While more testing lies ahead before operational use, the early results point toward a future where communities might get clearer, earlier warnings about the storms that matter most.