As massive winter storm hits, NWS has been using new AI-powered forecasts

With a winter storm set to blanket a 2,000-mile swath of the United States with snow, sleet, and ice, government forecasters are incorporating new, AI-fueled tools as they track the system.

Last month, the National Oceanic and Atmospheric Administration announced the rollout of AI-driven weather models as its head, Neil Jacobs, acknowledged that the country has fallen behind European counterparts and, in some cases, tech companies in forecasting accuracy.

At the same time, the agency is partnering with private companies to enhance observations that are essential for improving predictions. Poor radar coverage, known as radar gaps, has long plagued the National Weather Service, endangering residents when there’s a failure to detect hazards like tornadoes, ice, or heavy snow.

Both developments — the use of AI models and expanding observations — are part of the government’s push to fill those gaps and modernize its weather forecasting as events like this weekend’s storm underline increasing weather extremes.

“Getting the U.S. back to number one in weather forecasting is my top priority,” Jacobs said in an interview with the Weather Channel that aired in December. He added that AI-fueled developments in weather models were astounding. “I’m just starting to see stuff that blows my mind,” said Jacobs.

There are three new AI-driven products: One shows AI’s prediction of what will happen, another shows other scenarios that might play out, and the last — a hybrid — combines the AI prediction with NOAA’s ensemble numerical predictions.

Forecasters can access these whenever they want to add what Jacob Carley, a physical scientist at NOAA’s environmental modeling center, called “another tool in the forecaster’s toolbox.”

Snowstorms in particular rely on probabilistic forecasts to understand whether freezing rain or snow is more likely, said Carl Schreck, the interim associate director for the North Carolina Institute for Climate Studies at North Carolina State University. That means, for example, if you have a 75% chance of sleet, that percentage becomes more certain when you get that result 300 out of 400 times, versus 30 out of 40 times.

“Having lots of runs of the same model, there’s a better chance of seeing the whole range of options,” he said.

Ahead of the snowstorm, it was comforting to know models like these were being incorporated into forecasting, said Schreck, who spoke from Asheville, which, like dozens of cities, was bracing for severe impacts from the storm. He added the new technology marks the “most exciting time” in his 25-year career as a climate scientist.

Artificial intelligence entered the world of weather prediction a few years ago, but the suite of new forecasting models, along with government partnerships with private tech companies that are striving to fill the long-standing gaps in NOAA’s abilities, are part of what scientists and companies have hailed as a new era amid increasing demand for better predictions.

“People, communities, and institutions, including economic sectors, have become more attuned to the impact of weather on their safety and security,” said Rick Spinrad, who served as NOAA’s administrator under President Joe Biden. “So they are seeking more accurate forecasts with longer lead times.”

The modern weather forecast, developed in the early 20th century, relies on complex physical equations and the laws of thermodynamics to predict how the planet’s atmosphere will behave. It requires a lot of computing power and time to produce. Weather balloons, satellites, and buoys all gather data to add to these assessments.

But the past several years have seen the advent of a faster AI-fueled mode of forecasting that uses current data and is trained to analyze past weather patterns to predict future ones at times as accurately — if not more accurately — than conventional models.

That was on stark display this Atlantic hurricane season, when Google’s DeepMind AI model outperformed other traditional models in predicting both the path and intensity of the squalls.

Now Google’s code, which was further developed by NOAA, is helping to fuel the agency’s suite of AI-powered products that officially launched in mid-December, yielding improvements in performance that would normally take between five and 10 years to achieve with numerical weather predictions, said Carley.

“There’s been incremental improvements during that time, but nothing like what we’re seeing in the last couple of years with the AI revolution,” Schreck said.

While the technology will transform forecasting, he doesn’t believe it will change the need for forecasters.

“You still need the human forecaster to digest all of these models,” Schreck said. “It’s not producing the real forecast that a human would make right now.”

To make the best forecast, you need good weather observations — the surveillance and detection of weather patterns. Better observations can improve the AI-run forecasts, too. Now, private companies across the country are partnering with NOAA to use emerging technology to refine observations and predictions.

“This is a renaissance in terms of the weather ecosystem,” said Chris Goode, the founder and CEO of Climavision, which is working with the government to fill the blind spots, or radar gaps, in weather coverage across the country with lower level atmospheric surveillance.

Goode’s company joins several others using technology to bolster weather forecasts.

“The fact is that weather risk has outgrown the infrastructure that we built decades ago,” he said. The company affixes their radar to tall buildings or water towers in parts of the U.S. where there are known gaps, allowing them to pick up weather that might not be seen from the government’s higher vantage points or in places where there are no government radars installed.

The technology can be helpful in snowstorms like the one hammering the southeast this week, Goode said.

“Snowfall often produces weaker radar signals, occurs closer to the surface, and can organize into narrow, fast-moving bands that change conditions dramatically over short distances,” he said. His company’s radars can be especially useful in these situations. Because they operate “closer to the ground and between existing radar sites, they help resolve features like snow squalls and narrow lake-effect bands that can be partially missed or overshot by distant systems.”

Spinrad cheered the growth in commercial weather enterprises “for any number of reasons, not the least of which is that private sector now is employing many of the most capable meteorologists who were fired from the public sector just a year ago.”

The government lost hundreds of Weather Service employees to firings, resignations, and retirements amid the Trump administration’s efforts to trim the federal government. NOAA has since pledged to refill many of those jobs.

The radar gap is a long-acknowledged issue at NOAA, Spinrad said, but he worried about relying on a private company for a solution to operational problems. What happens if the sole provider loses funding or decides to change its product?

Regardless, Goode said, increasingly extreme weather was fueling a demand for his vision unlike anything he’d seen over 35 years in the industry.

“If we as a company attempted to do this a decade ago, it wouldn’t have been met with the same enthusiasm,” he said, “in part because we have seen this steady rise in volatile weather.”

And despite efforts to make forecasting more accurate, most people don’t have access to the watches, advisories, and warnings that come from scientists poring over these models and deciding what the public needs to know.

About 40% of countries lack multi-hazard early warning systems, according to the United Nations, and many of the existing systems are not equipped for the intensifying threats of climate change, including extreme heat, wildfires, and glacial flooding.

The AI weather revolution, Schreck said, could be a democratizing one. Fast, cheap AI models can be run from laptops, in some cases.

“It would be much easier to run specialized models for a developing country that may not have the infrastructure to run those models themselves,” he said.