“Deep Learning for phasing segmented telescopes.” by Maxime Dumont
The presentation will be held in English, and you can find the abstract below:
“Achieving both high angular resolution and frequent revisit times for observing Earth or other planets from low orbit presents several significant challenges. Balancing the need for a larger aperture with the associated increased costs and constraints requires innovative solutions. AZIMOV, a prototype of a segmented deployable telescope payload on a 6U CubeSat currently under development, addresses this challenge by using a telescope with a 30 cm aperture. This large primary mirror allows for a ground sampling distance of 1 meter in the visible spectrum. However, for optimal performance, precise alignment (or phasing) of the primary mirror segments is crucial. Due to CubeSat constraints, such as limited volume, power, and computational resources, conventional wavefront sensing methods are unsuitable. Focal plane wavefront sensing is the only practical approach for these small platforms, but traditional techniques are often iterative and computationally demanding due to the nonlinear relationship between the phase of the electromagnetic field and image intensity.
This thesis explores the use of deep learning to correct piston and tip-tilt aberrations on the four segments of the primary mirror from a single focal plane image. We demonstrate that our approach, using convolutional neural networks, can achieve near-diffraction-limited performance when observing a point source. This deep learning method is robust against noise and higher-order aberrations and surpasses traditional iterative methods in speed, accuracy, and robustness. Tests on experimental data confirm the adaptability of neural networks between simulated and experimental data and demonstrate the feasibility of our method on real data. For Earth observation from low orbit, or imaging unknown extended objects on Earth’s surface, our approach, refined for extended object analysis, shows performance approaching the diffraction limit, while highlighting limitations of single-image focal plane wavefront analysis due to characteristics of the observed object.”