WindBorne Breaks New AI Forecasting Records with “WeatherMesh-2” Model, Surpassing AI and Physics Gold Standards

‍

WindBorne has once again surpassed gold-standard physics- and AI-based weather forecasting models with a series of new accuracy gains by the latest version of its AI forecasting model, “WeatherMesh-2,” or “WM-2.”

WM-2 is a global, medium-range AI-based weather model that predicts core forecasting metrics across the Earth’s surface and atmosphere, including geopotential, wind speed and direction, temperature, precipitation, cloud cover, solar radiation, pressure, and humidity. The model predicts these metrics about 8% to 24% more accurately than the Global Forecasting Service (GFS), the European Center for Medium Range Weather Forecasting (ECMWF)’s HRES, and Google DeepMind’s GraphCast over equivalent time horizons; in many cases, WM-2 maintains this accuracy lead over other gold standard models even when predicting further into the future.

WM-2 forecasts 0.25 degree global grid forecasts for 14+ days at both hourly and six-hourly intervals. WM-2 currently generates forecasts four times per day, incorporating live real-time estimates of current weather analyses from large physics-based models, along with observations collected from WindBorne’s own atmospheric balloons.

Sample March-May 2024 WM-2 performance results:

• For 2 meter temperature, WM-2 is 14% more accurate than GraphCast at 14 days, 19% more accurate than HRES (10 days), and 23% more accurate than GFS (14 days).
• For 500mb geopotential, WM-2 is 8% more accurate than GraphCast at 14 days, 13% more accurate than the ECMWF’s’ HRES at 10 days, and 19% more accurate than GFS at 14 days.
• For 10 meter winds, WM-2 is 8% more accurate than GraphCast, 18% more accurate than HRES (10 days), and 21% more accurate than GFS (14 days)
• Across headline variables and lead times, WM-2 outperforms or performs as well as Microsoft’s Aurora on 79% of the targets evaluated while being one-quarter of the size by parameter count and nearly 30x faster in inference time.

Atmospheric Data and AI Modeling Break-Throughs

WM-2’s accuracy gains can be attributed to two unique advantages:

1. Data: WindBorne now incorporates its own real-time atmospheric data into WeatherMesh. This enables WM-2 to incorporate fuller, more accurate representations of weather snapshots in a process known as data assimilation. WindBorne gathers this critical data through its global constellation of long-duration, self-navigating “Global Sounding Balloons” (GSBs). GSBs can gather up to 50+ atmospheric profiles in a single flight and fly from 0 to 20 km in altitude, even triple circumnavigating the globe.
2. Modeling innovation: WindBorne’s world-class machine learning team continues to innovate fundamental machine learning methods to improve performance gains. The company’s latest breakthroughs maximize the time the model spends in what’s known as “latent space,” the abstract model-internal representation of weather over the planet which the model evolves forwards in time to make predictions. This means WindBorne’s model can predict longer time-horizon forecasts with significantly more accuracy; it avoids an encode-decode error at each timestep.
   1. While other models learn to predict only a single fixed timestep into the future, which limits their effectiveness at flexibly and accurately forecasting longer time horizons, WM-2 is tasked with accurately forecasting a range of timesteps during training. This innovation enables WM-2 to operationally predict longer time-horizons with more accuracy. 
   2. This innovation is only possible due to some fundamental training architecture breakthroughs that allow the model to leverage CPU RAM at the same time as GPU RAM during training. In order to do training in the model latent space for a six-day forecast, the training infrastructure has to keep over 200 unique instances of a latent-space-representation of weather conditions over the globe in memory at the same time per GPU. This is hundreds of gigabytes of data, which couldn’t fit in any GPU, even an NVIDIA H100, which has 80 GiB of VRAM.

WindBorne’s announcement comes just 10 months after the company unveiled its first groundbreaking AI-forecast model, “WeatherMesh,” in February 2024.

For a deeper look at WindBorne’s model performance across key weather variables, see WindBorne’s latest WM-2 case studies.

International Growth

WindBorne’s customer base and operations have continued to grow internationally since the company first announced WeatherMesh earlier this year. WindBorne began selling its atmospheric observational data on an operational basis to the National Oceanic and Atmospheric Administration (NOAA) two years ago, and now sells its data in both the public and private sectors. WindBorne has also started the first commercial pilots of its weather forecasting model WM-2 with partnerships in Asia and the U.S.

WindBorne’s autonomous atmospheric balloon constellation is powered by more than half a dozen international balloon launch sites, which include Palo Alto, Calif., upstate New York, and South Korea. The company has also established permanent sites in particularly under-observed regions such as sub-Saharan Africa and in Southeast Asia.

New high-resolution offering 

WindBorne’s revamped, WM-2 model now offers high-resolution local forecasts anywhere in the world for plots as small as 1 km by 1 km. WindBorne achieves this level of precision through a forecasting method known as “downscaling,” in which WM-2 works in tandem with an additional specialized WindBorne model tuned to generate high-resolution forecasts.

Record-breaking speed

WM-2, which is used in live, operational environments, is able to produce complete 10-day forecasts in nine seconds and complete 14-day forecasts in 13 seconds, about 143,000 times faster than the gold standard global weather model, the ECMWF’s HRES.

This output speed is significant for public sector forecast guidance and a wide range of commercial use cases, including energy grids, energy markets, aviation, live events, and more.

WindBorne’s machine learning and AI team, which has trained in world-class teams such as Fei-Fei Li’s Stanford University Lab and the University of California, Berkeley, has grown multi-fold since WeatherMesh’s debut. In the past year, the team has expanded from the company founding team to now making up nearly a fifth of WindBorne’s 40-person full-time team.

WindBorne continues to advance its forecasting technology and scale its global data collection. Over the next year, WindBorne plans to scale the size of its forecast models, increasing the number of parameters, which enables the model to understand more nuanced relationships in the data. WindBorne will also continue expanding its datasets, both through its own atmospheric data collection, as well as through supplemental environmental data. The team will continue to invest in underlying modeling techniques to improve forecast resolution, minimizing blur and other erroneous artifacts that can be generated through deep learning prediction techniques to support even longer time-horizon lead times.