Driven by the need of always more accurate models, space optics instrument-based observations push constantly towards high accuracy measurements that require an excellent knowledge of the instrument. To achieve this, current classical technologies are limited by the complexity of current instruments, calling for disruptive technologies to take over. Therefore, Airbus is currently integrating Artificial Intelligence (AI), responding to the call for new concepts. Here Airbus takes benefit of deep learning to detect complex patterns that would otherwise be impossible to properly characterize classically, opening the door for completely novel characterization paradigms and enabling manifold accuracy improvements. This work first focuses on obtained results on the detection of random telegraph signals (RTS) of CCD detectors under tests. By training a convolutional neural network (CNN) with RTS data, it has been possible to setup an algorithm achieving 20x faster data processing while increasing accuracy, providing unprecedented fast and performant RTS characterization. In another domain, multi-reflection-induced ghost stray light have been also characterized using CNN. Here, Airbus uses simulated data from optical software to generate 2D ghost maps used to train an algorithm capable of segmenting individual patterns. We show in this work that the appropriate architecture with optimized hyper-parameters achieves 97% accuracy. These ground-breaking results pave the way for a complete characterization of optical instrument ghosts that were so far neglected because of their complexity. It hence enables in the future more performant straylight correction algorithms as well as providing extended freedom in the design of space optical instruments.

No abstract

No abstract