WeatherMesh: Adapting Transformers for Weather Prediction

Deep learning models have emerged as powerful alternatives to traditional NWP methods, generating forecasts orders of magnitude faster. These models are trained on historical reanalysis data without explicitly solving for the complex physical equations that NWP models use, and are increasingly accurate at capturing the complex, multi-scale dynamics of the atmosphere.

We present WeatherMesh, an AI weather forecasting system developed by WindBorne Systems, to provide accurate weather forecasts for consumer decision-making processes. WeatherMesh has been running in real-time, four times per day, since March of 2024. During this time, the WeatherMesh forecast has been better than the IFS 24h forecast for 500 hPa geopotential by 14% on RMSE, and has consistently lower error on variables like 500mb wind speeds and 2m temperature, even out to 7 days. Given this success, WeatherMesh is actually being used in WindBorne’s operations to plan and guide flight trajectories for WindBorne’s constellation of Global Sounding Balloons.

Weathermesh uses a transformer-based architecture, which is well-suited to capture both long-range dependencies and fine-grained local patterns in atmospheric data. Transformers have also been used to great success in Large Language Models and computer vision, and we’ve built additional earth-system customizations on this framework.

WeatherMesh is trained on ERA5 data with a 0.25 degree resolution grid, and operationally it takes in IFS and GFS forecasts to generate forecasts with hourly output out to 7 days and 6-hourly output out to 21 days. Additionally, it uses perturbed initial conditions from ECMWF to generate ensemble forecasts which quantify forecast uncertainty and enhance reliability, especially in highly dynamic or uncertain atmospheric situations. WeatherMesh is also separately able to provide point forecasts for localized weather conditions at much higher spatial resolution, which is another necessity for more accurate and actionable decision-making in important sectors like agriculture, energy and emergency response.

As an example of this actionable decision making, WeatherMesh has also produced very impressive forecasts for tropical cyclone tracks. In order to assess the forecast skill of tropical cyclone tracks, we ran WeatherMesh retrospectively during the 2022 season, and found that for the 8 strongest tropical cyclones in the Northern Hemisphere, WeatherMesh forecast tracks were 40% better than GFS forecast tracks. Further, for Hurricane Ian, landfall location for a 72 hour forecast was improved by 200km, predicting the location of Ian's landfall within 10's of kilometers.