Machine Learning for Aero-Optical Wavefront Characterization and Forecasting

Abstract

The laser is a masterwork of the previous century of physics, but to harness its power coherently in the turbulent wilds of the sky remains a challenge. Free-space lasing hosts myriad applications, from direct energy transmissions for defense to secure communication channels that robustly support quantum entanglement. Aero-optics is the multi-disciplinary field underpinning such optical transmissions through atmosphere and fluid flows. Turbulent variations in the index of refraction, like those forming around boundary layers of airborne optical platforms, manifest aberrated wavefronts. Forecasting these rapid phase distortions allows us to rectify laser transmissions via adaptive optic engineering. Doing so hinges on the development of low-latency predictive techniques, for which we turn to advancements in data-driven algorithms and deep learning. This thesis thus introduces key concepts of aero-optics, wavefront sensing, and machine learning for data-driven physics. We then demonstrate three machine learning methodologies - optimized Dynamic Mode Decomposition, sensor fusion through shallow decoder networks, and forecasting via recurrent neural networks with shallow decoder outputs — for robust aero-optical wavefront sensing.