AI Enhanced Coherent Wavefront Sensing for ISR and Directed Energy Applications

Focusing light that is either received or transmitted through deep atmospheric turbulence is a key technological barrier to applications in ISR (intelligence, surveillance and reconnaissance) and DE (directed energy). In ISR, the goal is to image a distant target through deep atmospheric turbulence, and in DE the goal is to pre-correct the optical wavefront of outgoing light so that it focuses on a distant target. Both applications require high-throughput and high-accuracy estimation of the beam propagation model (BPM) in order to make phase corrections. Digital holography (DH) technology is the most promising technology for addressing this challenge because it allows for measurement of the optical phase and magnitude through the use of coherent illumination. This research proposes to develop a novel framework for high-accuracy, high-throughput and low-latency estimation of the BPM by combining state-of-the-art single-shot algorithms for phase distortion recovery known as DH-MBIR (digital-holography model-based iterative reconstruction) with state-of-the-art AI methods such as deep convolutional neural networks (CNN). By combining advanced physics models with empirical AI models of both the target image characteristics and deep atmospheric turbulence, we propose to dramatically improve accuracy of BPM estimation in difficult ISR and DE scenarios. The proposed framework, known as multi-agent consensus equilibrium (MACE), balances the requirements of fitting sensor data (i.e., the sensor agent) with fitting data models of the target (i.e., image prior agent) and the BPM model (i.e., BPM prior agent). The solution can be computed using an iterative algorithm for computing a fixed-point solution to an equilibrium condition. Basic versions of the MACE algorithm have been shown to result in substantial improvement in both image quality and BPM estimation accuracy. In the research, we propose to a) extend these methods to include advanced CNN based prior models of deep turbulence BPMs, b) incorporate ML surrogate models to reduce computation latency and improve throughput, and c) develop analytical and computational methods for concentrating energy on target for the DE application with deep turbulence. In order to fit our models to real deep turbulence characteristics that may be present in Navy hypersonic applications, we propose to make Phase I benchtop measurements of BPM characteristics in collaboration with JASR Systems and the USAF Academy. We also propose to design Phase II in situ measurements of BPM characteristics that can be used to quantify deep turbulence models and verify algorithm performance.