Assimilation of Long Duration Stratospheric Balloon Drift and Soundings in AI Weather Foundation Models

The rapid development of AI weather foundation models, such as ECMWF’s AIFS-ENS, promises to revolutionize operational forecasting by delivering competitive skill at a fraction of the computational cost of traditional numerical weather prediction (NWP). However, a critical gap remains: these models currently lack native, robust mechanisms for assimilating real-time, novel observation types, particularly in data-sparse regions. We present a preliminary framework for the first integration of ensemble data assimilation with AIFS-ENS.

This study uses a unique observational capability from Sorcerer long-duration stratospheric balloons, the only platform currently capable of providing simultaneous, high-frequency vertical soundings and multi-day Lagrangian float trajectories from the upper troposphere (12–14 km). To quantify the unique value of this multi-modal data, we conduct a series of Observing System Experiments (OSEs) assimilating: (1) vertical "yo-yo" profiles only, (2) Lagrangian drift velocities only, and (3) a combined hybrid dataset.

We investigate the hypothesis that while vertical soundings constrain thermodynamic profiles, the assimilation of continuous Lagrangian drift data provides a superior constraint on the upper-level wind field and jet stream positioning. We present an assessment of the technical feasibility of assimilating these diverse geometries into an AI-based background and offer a preliminary evaluation of their relative impact on forecast spread and error reduction. This work represents a novel step toward "observation-adaptive" AI prediction, exploring how next-generation hardware and machine learning models can be coupled to close global observing gaps.