This AI weather startup is beating government agencies at forecasting

A new AI weather forecasting tool was released today by the startup Wind systems It delivers more repeatable and accurate predictions on key variables than the world-leading system developed by European governments, thanks to advances in how sensor readings are fed into deep learning models.

Founded by A A group of Stanford University students In 2019, Windborne began building a better weather balloon, with the idea of ​​selling weather data. But with the arrival of deep learning models for weather forecasting in 2022, the team realized they could get even more value by building their own model, too.

Today marks the launch of the sixth version of that model, WeatherMesh, which the company says is more accurate than traditional and AI-driven forecasts produced by the European Center for Medium-Range Weather Forecast (ECMWF), the European government organization that meteorologists consider the leading provider of accurate weather forecasting today.

Windbourne says the new version of its model provides more accurate forecasts than traditional ECMWF and AI systems across several variables. One simple way to understand this, says Kay Marshland, Windborne’s chief product officer, is that WeatherMesh 6 is “as accurate after five days as traditional forecasts the day before,” especially with respect to surface temperature measurements.

WeatherMesh 6 produces forecasts every hour, instead of every six hours, as traditional models do. This resolution has now dropped to 3 km in Europe and the continental United States, where data quality is highest.

Traditional weather forecasts are generated through complex physical models that require expensive supercomputers to run, and take a long time to do so. AI models — built by startups and major labs like Google DeepMind — tend to move faster than physical models, but currently they don’t have as high an accuracy of as many variables and/or predict accurately over longer time periods.

However, weather AI is improving rapidly and is already being used in major government agencies around the world. Researchers are working to integrate it into systems used to collect weather data and produce general forecasts.

Windborne takes advantage of its unique combination of model building and data collection. The company now has about 400 balloons collecting sensor readings in flight at any given time, launched from 15 locations around the world. Advances in its current model come from improvements in how data collected by balloons is entered into forms.

“I personally don’t understand the business model of being an AI weather company without the advantage of a data set,” Windborne CEO John Dean told TechCrunch.

ECMWF’s superiority is due to the organization’s skills in “data assimilation” – working to transform disparate sensor readings into a comprehensive, machine-readable picture of the world. Currently, AI weather models rely on datasets produced by ECMWF and the US National Oceanic and Atmospheric Administration.

But Windborne and other organizations are working to feed data directly into models, and the company’s head of AI, Joan Creus-Costa, says that direct ingestion of data from its balloons and other sources is the main reason for the improvement in the new version of WeatherMesh. It took a year of tuning and redesigning the transformer-based model for the model to be able to deliver these predictions without losing stability.

“When we started (collecting data) we were still very dependent on ECMWF,” Dean said. “Today I expect that if we removed the initial conditions of the ECMWF, we would actually still achieve good results.”

The company suffered a scare last year when a United Airlines plane hit one of its balloons. While the plane sustained minor damage, no one was hurt, in part because Windborne followed US regulations on the size of its sensor array. However, the company has now added transponders to its balloons that report their location through the global flight control system, ADS-B, in an attempt to reduce the chances of another accident.

Windborne, which has raised $25 million in venture funding at a reported valuation of $85 million in 2024, is selling its balloon data to the National Oceanic and Atmospheric Administration (NOAA), where it is used by the U.S. Weather Forecasting Corporation, the U.S. Air Force and the U.S. Navy. The company also sells its forecasts to investors and commodity traders, but Dean says the company remains focused on building its model and data infrastructure on trading products, in part because of the changing nature of the information environment.

“I’m not trying to invest a huge team in building a SaaS product, if two years from now the way people want consumer information is through a proxy, right?” Dean said.