Simulation and Analysis of Turbulence Profiling Neural Networks for Tomographic Layer Reconstruction and Wide Field Image Correction in Telescopes

Abstract

Light that propagates through the atmosphere is subject to phase perturbations at layers of turbulent flow. For decades, traditional adaptive optics (AO) has used a deformable mirror (DM) to correct the phase at the system pupil. Since the correction is applied at the pupil – not at the layers of turbulence – traditional AO is only valid over a field of view of a few arcseconds for visible light. To obtain wide field image correction, the phase has to be compensated for at optical conjugates to the layers themselves. Doing so with traditional AO hardware increases system cost and complexity because multiple DMs are required. This has motivated the exploration of a software-based image correction technique called multi-object image correction (MOIC). As the cost of computational power continues to improve, MOIC has the potential to become a viable option for wide field turbulence compensation. Thus, the development of simulations and algorithms at the current moment will enable software-based MOIC for a range of applications in the future. In this work, a simulation of a MOIC system is built. The simulation enables exploration of machine learning based turbulence profiling and offline tomographic layer reconstruction to further wide field image correction without a DM. Chapter 1 provides background on turbulence, AO, MOIC, and machine learning. Chapter 2 describes how Shack-Hartmann wavefront sensor data can be used to measure the position of turbulence in front of the optical system. In Chapter 3, a simulation environment of an imaging system looking through a dynamic atmosphere is developed. We then present the layer signal to noise ratio (SNR) in Chapter 4 and demonstrate that it is a statistically valid metric for quantifying how difficult a layer of turbulence is to find in the atmosphere. Chapter 5 then details a process for using the layer SNR and the simulation environment to condition and generate large sets of data for training turbulence profiling neural networks. In Chapter 6, multi-layer turbulence reconstruction from a known turbulence profile is explored using different decomposition models. A four layer atmosphere is then modeled, measured, reconstructed, and compensated for to demonstrate end-to-end MOIC. Chapter 7 summarizes the work and suggests future areas of research.