WindBorne’s WeatherMesh system takes advantage of two technology trends: The rapid evolution of AI algorithms and the precipitous decline in the cost and size of computer hardware and wireless equipment.
WindBorne’s weather balloons, which cost about as much to manufacture as an inexpensive mobile phone, can orbit around the earth for weeks, using AI to precisely control their paths. WindBorne says it already employs the world’s largest constellation of weather balloons, and will increase it 100-fold to 10,000 balloons, giving a tiny startup as much visibility into the earth’s weather systems as heavily funded government agencies.
“As far as I’m aware, this makes us the first company to apply AI-based weather forecasting at scale in the real world,” said WindBorne co-founder Kai Marshland, who said the company markets its forecasts to a wide variety of potential customers, from maritime shipping companies to energy traders. The technology could also provide valuable data to climate researchers and help businesses save fuel, thereby reducing emissions.
Traditional computer-based weather forecasts are based on physics models that create atmospheric simulations. They require massive amounts of compute power, and are much slower compared to newer methods that utilize a different approach: deep learning.
Instead of using physics to understand what is happening in the atmosphere, the deep learning technique takes in vast amounts of data, from wind speed to barometric pressure, and picks up on patterns and cues that would be impossible for a human to find. Once it’s been sufficiently trained, the model can look at real-time weather data and predict where those metrics are likely to go, even without any prior knowledge of physics.
In November, DeepMind announced that GraphCast, a model trained on 40 years of weather data, was more accurate than the European Centre for Medium-Range Weather Forecasts, which is known as the gold standard in weather modeling.
WindBorne says its WeatherMesh model is 11% more accurate than DeepMind’s in the key forecasting metrics.
Another benefit of using deep learning instead of physics models is that although they require a lot of compute power for the initial training of the model, they are relatively fast and cheap to run after that initial process. DeepMind’s GraphCast and WindBorne’s WeatherMesh can both run on consumer-grade computers, and can complete in a matter of seconds work that would usually require a supercomputer under traditional models.
WindBorne is using a transformer model, a deep learning technique pioneered by Google and used in large language models like ChatGPT, which use next-word prediction to complete a sentence. WindBorne uses it to project what comes next in a weather forecast.
The company says it plans to combine both deep learning and the physics model approach to gain an advantage. Using proprietary data gathered by its balloons, it plans to create its own physics models for certain regions and then use the output of those models as data to pre-train its deep-learning model creating more detailed weather forecasts over certain geographies.
Over the weekend, WindBorne conducted a simulation of how its WeatherMesh model would perform in predicting the path of Hurricane Ian, a Category 5 storm that wreaked havoc in the Southeastern U.S. in 2022. The model followed the storm so perfectly that employees at first thought something must have been off, like the hurricane itself made it into the training dataset, giving it an unfair advantage. But the numbers checked out; the traditional models were off by around 200 kilometers.
