AI Couldn't Forecast the Texas Floods

Scientific American

3 days ago

CLIMATEWIRE | Artificial intelligence is showing promise when it comes to weather forecasting, but it still couldn't predict the Texas floods.
The best-performing weather models during the July 4 floods were traditional ones specially designed to produce local forecasts at high resolution. Global-scale models were far less accurate — and so were AI models, weather experts say.
'All those new fancy AI models? They missed it too,' said Daniel Swain, a climate scientist at the California Institute for Water Resources, in a live YouTube talk on July 7.
On supporting science journalism
If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
Some meteorologists say that could change. AI weather models are starting to exhibit an ability for deep learning of atmospheric physics, which means they could be capable of forecasting unprecedented weather events based on atmospheric conditions.
New AI models are "certainly capable of predicting 'out-of-sample' events — events that they haven't seen before,' said Corey Potvin, a scientist at NOAA's National Severe Storms Laboratory in Norman, Oklahoma.
But looming budget cuts at NOAA — along with years of lagging federal investment in AI weather systems — are a major hurdle for the improvement of federal AI weather models, experts say. It's the latest example of how President Donald Trump's efforts to shrink government could hobble the country's weather forecasting capabilities, at a time when extreme weather is on the rise.
Kim Doster, NOAA's director of communications, said in an email that budget cuts would not negatively impact the agency's research and forecasting priorities.
Commerce Secretary Howard Lutnick, head of the department that houses NOAA, "is committed to integrating advanced technologies like AI to deliver the fastest, most accurate data to Americans," she said. "This administration is working hand-in-hand with meteorologists and scientists to modernize research, cut forecast lead times, improve mapping, and streamline operations across the bureau."
Today's most advanced AI weather prediction models largely exist in the private sector. Many of them failed to see the Texas floods coming with the same accuracy as the high-resolution traditional forecasts.
One major reason is that many AI models are still focused on forecasting large-scale weather patterns at the global level, according to Russ Schumacher, a meteorologist at Colorado State University and Colorado's state climatologist.
'Forecasting precipitation at the local scale is very challenging, and has not really been the focus of most of the AI models in use now,' he said in an email.
That's despite some recent suggestions that the Texas forecasts could have benefited from more investment in AI prediction at the National Weather Service. Tim Gallaudet, who served as acting NOAA administrator during the first Trump administration, suggested in a July 7 op-ed that NWS should 'incorporate more artificial intelligence' into its atmospheric, oceanic and hydrologic modeling systems for more accurate forecasts during incidents like the Texas floods.
But some scientists have expressed concerns about AI's ability to forecast record-breaking weather events, like the extreme rainfall that triggered the Texas floods. AI systems are often trained on historical weather data, and extreme events are — by definition — rare. That means there aren't many examples of them for AI systems to learn from.
In a 2023 comment published in the scientific journal Nature, weather experts Imme Ebert-Uphoff and Kyle Hilburn warned that AI systems are 'often unpredictable when the program operates under conditions that it has never encountered before,' adding that extreme weather events 'might therefore trigger highly erratic predictions.'
Potvin predicted new AI models could forecast rare events, though not quite as accurately as they would if they had lots of examples to train on. And although most AI models are still focused on large-scale weather patterns, high-resolution models are likely on the horizon.
NOAA is working on some local weather-modeling projects that include AI components. The National Severe Storms Laboratory's experimental Warn-on Forecast system, or WoFS, is designed to rapidly incorporate radar and satellite observations into a high-resolution model. It can produce updated forecasts about every 15 minutes, increasing meteorologists' ability to accurately warn communities about sudden extreme events, like flash floods.
NSSL scientists are also perfecting an AI version of WoFS, known as WoFSCast. By design, it can only perform as well as the original non-AI model — but it can theoretically produce forecasts much faster and with far less computing power, making it a cheaper option for local NWS offices.
There's also NOAA's High-Resolution Rapid Refresh model, known for its ability to forecast storms at the local scale. HRRR was one of the models that best predicted the rainfall in Texas — and scientists are developing an AI version as well, a model known as HRRRCast.
'As far as I know, WoFSCastand HRRRCast are the only [AI] models currently being developed for higher resolution prediction,' Potvin said.
Lag in investment
NOAA still lags far behind the private sector when it comes to investment in AI weather prediction.
That's a big concern for NWS forecasts, some experts say. AI is swiftly becoming a new frontier in weather modeling, and it could easily become an asset to NWS meteorologists — if NOAA had more resources to invest.
'The private sector is well ahead of where NOAA is now, to the point that even if we were in normal budget cycles, I'm not sure they could catch up,' Mary Glackin, former president of the American Meteorological Society, said at an AMS-hosted panel last week.
Meanwhile, the White House has proposed around $2.2 billion in cuts to NOAA in its budget request for fiscal year 2026.
Chief among these is the elimination of NOAA's entire research arm. That includes the agency's large network of cooperative research institutes and laboratories, like the NSSL, where researchers are still improving forecasting systems like WoFS and its AI counterpart.
Scientists have warned that these cuts would damage NOAA's weather forecasting capabilities, putting communities at risk when extreme weather events strike.
The private sector alone can't make up for lagging federal investments in weather forecasting technology, scientists and meteorologists have warned. Agencies like NWS are invested in public service and free forecasts, with the aim of ensuring that all U.S. communities — even those with limited resources — have access to high-quality, life-saving weather warnings.
More public-private partnerships could help NOAA get a jump on AI weather system development, Glackin suggested. Such an arrangement 'meets the needs of the private sector, who are looking for a profit and a competitive edge, but remains true to the public service concept and not leaving the least behind,' she said.
But such partnerships require the continued existence of research infrastructure at NOAA — which might not survive if Congress follows through with Trump's proposed cuts.
Meanwhile, AI isn't the only frontier in weather forecasting. Traditional weather models also improve year over year as scientists collect and incorporate more data. That's how hurricane forecasts become so advanced over the last few decades.
"As big a fan as I am of AI, it would be a mistake to put all of our investment into AI and then neglect the traditional side of weather modeling,' Potvin said. 'Because that in the end, would be undercutting future AI development.'
Meteorologists have warned that traditional weather forecasts will plateau — or even degrade — if Trump's proposed cuts go into effect and hamstring NOAA's research capabilities.
'I worry about the loss of investments in science,' Brad Colman, another former AMS president, said at the July 10 panel. 'That's our seed corn, and the impact of that will be long-lasting. So I really hope that a greater wisdom will prevail, and that we will maintain that capacity.'