Flexible Data Assimilation with Conditional Diffusion

Forecast weather significantly influences naval operations throughout planning, logistics, and mission execution. Winds can severely complicate projectile trajectories and aerial operations, rain can impact electronic and mechanical equipment, visibility directly affects engagement decisions, to list a few considerations. Accurate, high-resolution forecasts are invaluable, and the ability to rapidly perform forecasts using localized data is even more so.ÝMachine learning-based weather prediction (MLWP) has reached the stage where it is technically feasible to utilize temporally/spatially sparse and irregular multi-source data to accurately generate forecast ensembles in real time. Skydwellerís solution is a state-of-the-science conditional diffusion network with observation data-guided generation. Along with a proven ML framework that can generate high-resolution ensembles in real time, Skydweller provides a unique data source: onboard multi-mode, AESA weather radar.ÝMoreover, the Skydweller platform itself is the largest and most mature solar aircraft in the world.ÝWith an endurance measured in weeks/months and a high weight and power payload capacity, the Skydweller aircraft will provide naval commanders with entirely new mission flexibility and opportunities. Skydweller capabilities will be critically enhanced by the integration of an onboard forecasting system. First, in a reduced communications scenario, the platform can use short-term forecasts based on data available from aircraft sensors to autonomously adjust its mission plans. Second, in typical operations it can host a robust communications payload and act as a communications hub, ingesting data from other naval entities in the domain and providing the forces with a high-quality forecast. Funding this effort will provide the Navy with a powerful MLWP forecasting system independent of the Skydweller platform, but also the option to utilize the ensemble forecaster within the aircraft along with the systemís unique capabilities, including the valuable data from the onboard weather radar.ÝThe conditional diffusion methodology detailed in this volume generates an ensemble of forecast atmospheric states over a region based on the previous states and observation data that may be temporally and spatially sparse and irregular. In the Phase I base, data assimilation will focus on individual observations of single variables. While evaluation will primarily take place over readily available weather station surface observations, aircraft telemetry data from past flights will also be utilized. The Phase I option will advance to more complex data, such as weather radar (both surface and onboard), satellite imagery, and qualitative information.