Andre Fogal

Topic

Low-order Wavefront Sensing Through Machine Learning: From the SPIDERS Bench to On-sky Observations

Department of Physics and Astronomy

Date & location

• Monday, August 10, 2026
• 2:00 P.M.
• Clearihue Building, Room B007

Examining Committee

Supervisory Committee

• Dr. Christian Marois, Department of Physics and Astronomy, University of Victoria (Co-Supervisor)
• Dr. Jon Willis, Department of Physics and Astronomy, UVic (Co-Supervisor)

External Examiner

• Dr. Peter Wizinowich, W.M. Keck Observatory, Hawaii

Chair of Oral Examination

• Dr. Simon Devereaux, Department of History, UVic

Abstract

Advancements in astronomical technologies over the past 20 years have led to the ability for astronomers to do direct imaging of exoplanets, in which light originating from planets around other stars is captured with ground- and space-based telescopes. The Subaru Pathfinder Instrument for Detecting Exoplanets and Retrieving Spectra (SPIDERS) is a pathfinder to test focal plane wavefront sensing technology for direct imaging of exoplanets. SPIDERS implements two kinds of wavefront sensors, the focal plane Self-Coherent Camera for removing speckles and the Lyot-based Low Order Wavefront Sensor (LLOWFS) for correcting low-order aberrations. In this thesis, I developed a machine learning-based wavefront reconstructor for the LLOWFS, demonstrating its massive performance increase over traditional linear methods, including a much larger linear response regime and strongly reduced modal crosstalk in reconstructed signals. I make use of a unique Julia package optimized for lightweight neural networks for science applications to avoid increasing computation time, allowing the machine learning reconstructor to function at speeds up to 1000 Hz. I show results from bench testing and from on-sky observations at the Subaru Telescope during the SPIDERS observing run which took place from 29 December 2025 to 2 January 2026. I perform spectral analysis on both bench and on-sky residuals, demonstrating a rejection of low order modes up to 50 dB. This work demonstrates the potential for performance gain from implementing machine learning reconstruction, without sacrificing loop speed.