Paper 14145-292
Paper 14145-292
Analysis of an augmented extended Kalman filter and convolutional neural network approach for detecting planet signals in dark zone maintenance datasets from the Decadal Survey testbed
9 July 2026 • 17:30 - 19:00 CEST | Room B4-M3
Abstract
Direct imaging of exoplanets requires starlight suppression, which can nominally be done using coronagraphs and deformable mirrors. However, starlight suppression is not perfect and residual speckles still challenge traditional post-processing methods. Recent work has shown that AI-based algorithms, particularly Convolutional Neural Networks (CNNs), can improve exoplanet detection. We extend prior efforts that combined a focal plane wavefront sensor (FPWFS) with an augmented EKF and CNN to detect very dim planets in conditions similar to the Habitable Worlds Observatory. The CNN learns to discriminate noise from injected planetary signals, leveraging drift information captured by the EKF, using the Decadal Survey Testbed. We present the preliminary results and compare performance against traditional high-contrast imaging methods.
Presenter
Celvi Lisy
Massachusetts Institute of Technology (United States)
Celvi Lisy graduated with a B.S. in Mechanical Engineering from Olin College of
Engineering in 2023 and a Masters in Aeronautics and Astronautics from Massachusetts Institute of Technology in 2025. She is currently pursuing a PhD in Aeronautics and Astronautics at MIT under Professor Kerri Cahoy as part of the STAR Lab. At Olin, Celvi was one of the founding members and ultimately the Lab Manager of the Olin Satellite + Spectrum Technology and Policy Group (OSSTP). She also served as the Structures Lead for the multi-institutional, NSF-funded SWARM-EX CubeSat mission. At MIT, she has worked on the CLICK B/C mission as a systems, mechanical, and operations engineer and has worked on the SAT grant. Past experience has included research at Los Alamos National Lab in vibration analysis, research in on-orbit servicing and post-mission disposal, and work on systems, ground stations, and structures subteams for small satellite missions that have led to thirteen publications and counting!