Using Deep Learning to Improve Euclid Data Analysis

The project focuses on developing new artificial intelligence methods to monitor and
correct radiation damage in space telescopes, enhancing our knowledge of the damage that
detectors experience in space. The Euclid mission, launched in July 2023, is an ESA space
mission with the objective of mapping the geometry of the Universe and better understand
dark matter and dark energy. While in space, Euclid will be bombarded with highly energetic
particles mainly from the Sun, which will slowly degrade the detectors and thus have a have
a major impact on the scientific data unless corrected for.
The Centre for Electronic Imaging (CEI) at the Open University has been involved in the
Euclid mission since its conception as part of the Euclid VIS instrument development work,
doing detector characterisation and leading the radiation damage testing. As part of this
work, the CEI has developed the Trap Pumping technique, which allows for the
characterisation of single defects in the silicon lattice of irradiation damage devices. As this
will allow for a much deeper understanding of the actual damage the devices is subject to
when in space, trap pumping has been adopted for Euclid VIS and will be performed on a
daily basis as part of the in-orbit calibrations. This will be the first-time trap pumping is done
in space.
The CEI is part of the radiation damage correction efforts, and we are working on
implementing our analysis routines into the automated data analysis pipeline for the Euclid
mission. Trap pumping is reasonably well understood, and the trap parameters can in
principle be found by simple curve fitting, however, there is a wealth of special cases, where
this approach breaks down. This is especially when there is a high level of noise that makes
the fit more uncertain or at high irradiation levels, where multiple traps can interfere with
each other. The Euclid VIS instrument consists of 36 detectors with a total of over 600
million pixels. Each detector will contain thousands of traps even before it is irradiated, and
this number is expected to rise ten or hundred fold during the mission. Automated methods
of trap detection and fitting is therefore absolutely essential. We thus plan to explore
machine learning methods to improve the detection and fitting of these special cases in
order to improve the overall precision of the obtained trap parameters. In particular, we will
study a variety of deep learning techniques for computer vision (e.g., convolutional neural
networks, vision transformers, capsule networks) and determine the best architecture in
this domain. This will help improve the quality of the irradiation damage correction for
Euclid VIS and increase our understanding of the damage detectors are subject to in space.