(Bloomberg Businessweek) -- With the pull of a cord, a balloon laden with sensors careens into the cloudless sky above a parking lot in Palo Alto, California. Below, three of the co-founders of WindBorne Systems Inc., an artificial intelligence weather-forecasting startup, crane their necks to watch what looks like a jiggling jellyfish begin a multiday journey across the US and potentially beyond.
The sensors on the balloon will monitor wind speed, temperature and atmospheric pressure as it travels through the atmosphere for as many as 12 days. It’s part of WindBorne’s multipronged strategy to make better weather predictions by collecting data, which it sells primarily to energy traders looking to get ahead of potential strain on the grid. Since 2021 the startup has also been working with the National Oceanic and Atmospheric Administration, which is testing whether the balloon data can improve its forecasting abilities.
In addition to that, WindBorne is now developing its own forecasting tool powered by AI. With what could be a record-shattering hurricane season, weather researchers are watching the company’s projections—and those of other emerging AI forecasters. “This is one of the first seasons to really test these [new models],” says Matthew Chantry, machine learning coordinator for the European Center for Medium-Range Weather Forecasts (ECMWF), an intergovernmental organization.
John Dean, Kai Marshland, Andrey Sushko and Joan Creus-Costa founded WindBorne in 2019. Self-avowed sci-fi nerds with backgrounds in engineering who share an obsession with Battlestar Galactica (at every launch they exclaim in unison, “So we say all,” a reference to the TV show), they were motivated largely by a love of releasing things into the sky. Weather balloons—which government agencies and even amateurs launch from hundreds of locations around the world each day—were a way to indulge that passion while improving the existing technology.
Since weather balloons first took flight in 1890, they’ve been subject to the vagaries of the atmosphere, riding the wind and going wherever it blows. WindBorne’s balloons still do this, but they also have the ability to navigate on their own, thanks to a unique ballast system and a suite of sensors about the size of a Starburst candy. The company can set a flight path at takeoff, telling the balloon to sample a vertical slice of the atmosphere at regular intervals, for example. But operators can also take control and modify the flight path in real time.
Regular launches from four locations in the US, South Korea and the Cabo Verde islands off Africa’s west coast provide a steady stream of global data, whereas specially programmed launches capture information about specific weather events. NOAA, for example, worked with WindBorne in 2022 to send balloons from South Korea and Hawaii across the Pacific in a bid to monitor atmospheric rivers that deluge the western US.
On its website, NOAA has called the data that WindBorne provides a “valuable addition” to weather balloon data. The balloons’ ability to navigate the atmosphere vertically means the startup can capture differences in temperature, pressure, wind direction and more at various heights. This fine-grained data gives forecasters a clearer picture of what’s going on in the atmosphere. If traditional weather balloon data is akin to trying to figure out how warm the ocean is by sticking a toe in, the startup has instead decided to dive all the way in.
So far, NOAA has tested whether incorporating WindBorne’s data into experimental models actually improves predictions. The results have been promising: Doing so would have improved the accuracy of the agency’s forecast for the track of 2022’s Hurricane Fiona by 18%, according to results presented at the American Meteorological Society conference in February.
NOAA and ECMWF both work with a number of weather startups. A spokesperson for ECMWF also noted that it still hasn’t fully vetted the impact of WindBorne’s data on its model.
WindBorne is using its balloons for its own AI-powered forecasting model. The system is trained on a vast dataset put together by ECMWF of hourly weather from around the world since 1940. Inputting recent data allows it to create a 10-day forecast of everything including temperature and the paths of storms. In traditional weather modeling, meteorologists process observations using physics equations; AI instead looks for patterns in past weather to predict the future, which can theoretically increase accuracy, ECMWF’s Chantry says.
WindBorne’s model isn’t available for public use, but the startup has published case studies showing the quality of its forecast by applying its model to past hurricanes. In a blog posted in May, the company compared its ability to predict eight cyclone tracks with NOAA’s, and its analysis showed that its model outperformed the agency’s physics-based tool. WindBorne also releases comparative analyses showing that since March its model has routinely been better at predicting near-future weather than the NOAA and ECMWF models. The company is in the process of assimilating the data from its weather balloons, which it expects will yield major improvements in accuracy.
WindBorne is not alone in applying AI to weather forecasting. In July 2023, researchers at Huawei Technologies Co. published a study in Nature showing they’d created an AI-based model that could predict the weather more accurately than ECMWF’s physics-based model, which is considered the world’s gold standard. A few months later, Google scientists published a paper showing that an AI-based model they’d built was even better. The papers focused on forecasting "at the global level using deep learning" and "they hit with a vengeance," sparking a surge of interest in AI weather modeling, says Amy McGovern, a machine learning and weather prediction researcher at the University of Oklahoma.
In February, WindBorne said its forecast bested Google’s in geopotential height, a key metric that helps meteorologists track weather systems, though its results have not been peer-reviewed. WindBorne’s Dean also acknowledges that it can be hard to nail down what exactly counts as accurate, because all three analyses use slightly different metrics. “Whenever someone’s making comparisons, there’s 1,000 ways that you can compare it,” he says.
This hurricane season will provide a true measuring stick for WindBorne. It’s already been marked by the strongest June storm on record, and the superheated oceans are primed to spin up more dangerous cyclones. Throughout the season, WindBorne is launching balloons from Cabo Verde, a hurricane formation hot spot. Dean says the startup has “more inbound interest than we can handle right now” in terms of customers.
Those customers will have other choices, too. Rival AI forecasting startups including Atmo, Jua and Tomorrow.io are all hoping to sell their weather predictions to government agencies and weather-dependent industries such as energy and aviation. ECMWF and NOAA are also working on AI models of their own, drawing on a wealth of in-house meteorology and modeling expertise.
Then there’s the possibility that an inaccurate AI forecast for a high-stakes event such as a hurricane or wildfire conditions could undercut public trust in the technology. Although AI models are showing great promise at predicting the paths of hurricanes, physical models have a clear lead for now at predicting intensity, says Chantry of ECMWF. “Neither system is perfect for anything.”
There are concerns, too, that AI models might not be able to capture extreme events such as Category 5 hurricanes, given their rarity. Climate change shifting global weather patterns could also make training data obsolete.
WindBorne’s founders are aware of these risks. But they also point to the promise of what a better-observed world and more powerful forecasts could mean for society. A 2024 National Bureau of Economic Research working paper found that hurricane forecast improvements since 2007 have reduced the costs of damage and loss of life by 19%, saving billions of dollars.
The long-term goal, Marshland says, is to have 10,000 balloons aloft at a time, collecting a wide range of data and bringing a greater understanding of what’s happening in the atmosphere. “What I want is for weather to be like a calendar,” he says. “It affects your operations, you know that it exists. But it isn’t something unexpected.”
(Updates paragraph 12 with quote from Amy McGovern. An earlier version corrected information about WindBorne’s founders.)
©2024 Bloomberg L.P.