US weather technology startup Windborne Systems has released the sixth version of its AI weather forecasting model, WeatherMesh. The company claims that the new model's predictions on several key variables have surpassed those of both the traditional and AI systems of the European Centre for Medium-Range Weather Forecasts (ECMWF), and its update frequency has increased from once every 6 hours to once per hour.
Founded in 2019 by a group of Stanford students, Windborne initially focused on developing weather balloons and selling observational data. With the rapid development of deep learning weather models after 2022, the company began to shift towards building its own forecasting models, hoping to combine data collection with model capabilities.
Forecast frequency is higher
WeatherMesh 6 has now improved its resolution to 3 kilometers in Europe and the continental United States, which is the region with the best data quality for the company. Kai Marshland, Chief Product Officer of Windborne, said that, taking surface temperature as an example, this version's prediction accuracy, given 5 days in advance, is approaching the level of traditional models that provide 1 day in advance.
Traditional weather forecasting relies on complex physical models, typically requiring expensive supercomputers and taking a long time to compute. In contrast, AI models generate results faster, but have historically lagged behind traditional systems in terms of resolution, variable coverage, and stability over longer time scales.
Building your own data network is key.
Windborne currently launches weather balloons at 15 locations worldwide, with approximately 400 balloons in the air collecting sensor data at any given time. The company states that the main improvement in this version is not just the model structure itself, but rather how to input data from balloons and other sources more directly into the model.
ECMWF has long been considered one of the world's leading weather forecasting organizations, largely due to its "data assimilation" capabilities—that is, integrating scattered sensor readings into a unified input that can be used by models. Currently, most AI weather models still rely on datasets compiled by ECMWF and the National Oceanic and Atmospheric Administration (NOAA).
Joan Creus-Costa, head of AI at Windborne, stated that the core improvement in WeatherMesh 6 lies in the company's direct input of its own balloon data and other observational data into the model. To achieve this, the team spent a year adjusting and refactoring the Transformer-based model to maintain stability while improving accuracy.
Clients include government agencies and trading institutions.
Windborne sells balloon data to NOAA, which is used in the U.S. weather forecasting system. Its clients also include the U.S. Air Force and Navy. In addition to government agencies, the company sells weather forecasts to investors and commodity traders.
However, the company's management stated that the current focus remains on continuing to build models and data infrastructure, rather than investing heavily in consumer-facing software products. They anticipate that information distribution methods may continue to evolve, and the form of commercial products may not necessarily remain within the traditional SaaS model.
Last year, Windborne drew attention after a balloon collided with a United Airlines plane. The incident resulted in only minor damage to the aircraft and no injuries. Since then, the company has added transponders to the balloons and uses the ADS-B global air traffic control system to transmit their positions, aiming to reduce the likelihood of similar incidents recurring.
Additional information:Public information shows that Windborne has raised a total of $25 million, and its valuation in 2024 was approximately $85 million.
