Assimilating GSB Observations into WeatherMesh AI model

Recent advances in artificial intelligence (AI) for medium-range weather forecasting have led to their growth in popularity. AI weather models rely on pre-existing reanalysis for training and operational datasets from weather centers for initialization. However, there has been limited work to develop and understand the performance of AI weather models using real-time data assimilation (DA) with direct observations. In this work, we address this need by exploring the effect of assimilating the atmospheric vertical profiles collected by WindBorne System’s Global Sounding Balloons (GSB) on the forecast skill of our developed AI model, WeatherMesh.

WindBorne Systems, a startup based out of Palo Alto, CA, previously developed GSB to enable novel observations from constellations of balloons traversing the troposphere for extended periods while collecting vertical atmospheric profiles of pressure, temperature, humidity, and wind. WeatherMesh is an AI weather forecasting system, also developed by WindBorne Systems, to provide accurate weather forecasts for consumer decision-making processes. Specifically, we use the Data Assimilation Research Testbed (DART) software to employ the ensemble Kalman filter (EnKF) for assimilating the GSB balloon data, then produce updated analyses for generating WeatherMesh forecasts. We initially determine the reliability of the ensemble and performance statistics of the resulting DA forecasts by comparing them to the Windborne GSB observations. We then compare the DA forecast performance to real-time deterministic WeatherMesh forecasts and cold-start WeatherMesh forecasts initialized from the Integrated Forecasting System (IFS) Ensemble analyses by the European Center for Medium-Range Weather Forecasts (ECMWF). The outcomes of this work will provide insights on data assimilation impacts within AI forecasting systems, like WeatherMesh.