WindBorne Systems: the AI that challenges weather forecasts

WindBorne Systems: A Technological Advancement in Weather Forecasting

The startup WindBorne Systems has just unveiled a new AI-based weather forecasting tool, promising more frequent and accurate predictions than those provided by government agencies. With significant advancements in integrating sensor readings into deep learning models, this technology could redefine industry standards.

Founded in 2019 by a group of Stanford students, WindBorne initially developed an innovative weather balloon aimed at selling weather data. However, the emergence of deep learning-based forecasting models in 2022 prompted the team to create its own model, seeing an opportunity to maximize the value of their data.

WeatherMesh 6: Unmatched Accuracy

According to WindBorne, their latest model, WeatherMesh 6, offers superior accuracy compared to traditional systems and the AI models of the ECMWF (European Centre for Medium-Range Weather Forecasts). Kai Marshland, product director at WindBorne, explains that WeatherMesh 6 can predict with the same accuracy five days in advance as a traditional forecast can the day before, particularly for surface temperatures.

This sixth version of the WeatherMesh model, launched today, is touted as being more accurate than ECMWF forecasts, a major player in the weather forecasting field. WeatherMesh 6 stands out for its ability to produce forecasts every hour, unlike traditional models that do so every six hours, and offers a resolution of 3 km in Europe and the continental United States.

AI Serving Meteorology

Traditional weather forecasts rely on complex physical models that require expensive supercomputers and a lot of time to produce results. AI-based models, developed by startups and labs like Google DeepMind, are generally faster, although their resolution is often lower for long-term forecasts.

However, AI in meteorology is progressing rapidly and is already being adopted by government agencies around the world. Researchers are striving to integrate this technology into existing systems to enhance the collection and production of public forecasts.

A Unique Data Collection Method

WindBorne distinguishes itself through its data collection method and the construction of its models. Currently, about 400 of the company's balloons are in flight, collecting data from 15 global sites. The improvements in their model stem from how these data are integrated into the learning models.

John Dean, CEO of WindBorne, emphasizes the importance of a data advantage for an AI-based weather company. The superiority of the ECMWF lies in its expertise in data assimilation, transforming disparate readings into a coherent and machine-readable image. For now, AI models depend on datasets from the ECMWF and NOAA.

Towards Data Autonomy

WindBorne and other organizations are working to feed their models directly with data. Joan Creus-Costa, head of AI at WindBorne, states that the direct ingestion of data from their balloons is crucial for improving WeatherMesh. It took a year to adjust the transformer-based model to provide stable forecasts without relying on the initial conditions from the ECMWF.

Dean anticipates that WindBorne could soon operate without the initial conditions from the ECMWF while maintaining high performance.

Safety and Expansion

Last year, an incident occurred when a United Airlines plane struck a WindBorne balloon. Although the plane sustained minor damage, no one was injured. Since then, the company has equipped its balloons with transponders signaling their position via the ADS-B system to prevent future accidents.

WindBorne has raised $25 million in venture capital funding, with an expected valuation of $85 million in 2024. The company sells its data to NOAA, as well as to the U.S. Air Force and Navy. It also offers its forecasts to commodity investors and traders, although its primary goal remains the development of its model and data infrastructure.

Dean expresses reluctance to heavily invest in a SaaS product, preferring to focus on evolving the information environment and how consumers access data.