This project will be jointly supervised with Tony Travouillon and Damien Gratadour from the research school of astronomy and astrophysics
Adaptive Optics is an instrumental technique for the correction of dynamically evolving aberrations in an optical system. Such an instrument is generally composed of one or several wavefront sensors (WFS) used to measure the distortions introduced by the turbulence, a RTC which using these measurements computes and sends commands to one or several deformable mirrors (DM) that compensate for these distortions before sending the light to the science instrument. This technology provides, on ground based astronomical telescopes, a significant improvement in resolution (up to a factor 10 for an 8m telescope and 50 for the future 30 to 40m telescopes), especially in the near-infrared by compensating, in real-time, for the effect of atmospheric turbulence.
While this method to compensate for distortions in real-time provides enhanced and more stable image quality out of the telescope, astronomers usually make use of post-processing techniques to extract more information out of the science data. A powerful approach is to leverage the AO system’s telemetry, which records the residual optical errors at any time during the exposure, in order to reconstruct a high fidelity estimate of the overall optical system’s impulse response. This calibrated Point Spread Function (PSF) estimate can then be used to invert the image formation equation and, since the telescope is acting as a low-pass filter, enhance numerically the contrast on the finest details in the final image.
Such technique has been developed and tested in the 90’s on astronomical telescopes with success on bright guide sources. It is based on careful modeling of the AO system error budget and is solved with a linear approach, sometimes regularized with some a priori on the turbulence statistics. However, when it comes to advanced AO concepts on the largest optical telescopes with fainter guide sources, the system performance can be severely impacted by an imperfect forward modeling of the turbulence statistics and noise, leading to the introduction of low accuracy a priori in the regularization scheme, as well as low accuracy in the calibration of the deviations of the instrument optical train from the model, which can evolve dynamically during the observations.
The objectives of this PhD project is thus twofold:
Develop and evaluate a deep learning architecture that robustly models a very high fidelity point spread function from adaptive optics telemetry, and the science path. We expect to leverage as a baseline existing linear models and existing studies of non-linear effects, and shall draw on a large corpus of both real-world and simulation data to progress this objective.
Evaluate the use of developped models and their associated architectures in end-to-end simulations of adaptive optics control.
Excellent understanding of Machine Learning, and in particular Deep Learning principles and practice.
Some knoweldge of optimised control under uncertainty in high-dimensional spaces would be a plus.
Be the best astronomer you can be.
point spread function