Vignesh Kumar Pandian Sathia

In astronomical imaging and adaptive optics (AO), when the atmosphere is modeled as a collection of wind-driven, frozen turbulent layers, it is possible to recover wavefront estimates that are more accurate than those obtained without accounting for temporal correlations. To make use of this frozen flow model, however, it is necessary to know the intrinsic details about turbulent layers, specifically how fast one is moving. Traditionally, this information comes from pupil plane wavefront sensor data, but that approach brings its own set of challenges, including susceptibility to noise and latency. Developing a method to estimate the characteristics of frozen turbulent layers, i.e., speed and direction, from focal plane images can improve restoration of archival speckle data lacking wavefront sensor input and reduce AO system latency by en-abling wavefront prediction from focal plane observations alone. Therefore, as an alternative to traditional methods, this work investigates a data-driven framework to predict wind profiles from the focal plane images themselves. We show the viability of using focal plane data, rather than relying on pupil plane measurements, by predicting the speed and direction of a single dominant atmospheric layer in synthetic data produced by a 1-layer turbulence model. Our predictive model design utilizes an autoencoder architecture augmented with VI CReg in the latent space. We demonstrate that the use of this regularization enhances robustness, and reduces the number of consecutive image frames needed for an accurate prediction.