Using AI to Enable Autonomous Exoplanet Direct Imaging
Name
page_pagec_phd_aeroastro_2026_thesis.pdf
Size
17.58 MB
Format
Adobe PDF
Checksum (MD5)
37b0ef500483e1584e1c371943546134
Author(s)
Page, Christine
Advisor(s)
Cahoy, Kerri
Fan, Chuchu
Gizzi, Evana
Oppenheimer, Rebecca
Hess, Herbert
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Directly imaging and detecting exoplanets with a high-contrast imaging instrument must overcome residual starlight and quasi-static wavefront errors which pose significant challenges for detection. In this work, we apply an Extended Kalman Filter (EKF) and two Artificial Intelligence (AI) models: a Convolutional Neural Network (CNN) and a Vision Transformer (ViT). We use the AI models to enhance exoplanet detection in high-contrast images. In the wavefront sensing and control system, the EKF estimates the real and imaginary components of the open-loop electric field as well as the incoherent intensity for each pixel. The EKF is assembled using cells of pixels in the dark zone where each cell is the size of a potential planet point spread function (PSF). This aids the estimator as the properties of the planet PSF pixels should be related. The cell mapping for the EKF is initialized randomly assuming no prior knowledge of the potential planet location. The AI models are trained and tested on images from the High-contrast imager for Complex Aperture Telescopes (HiCAT) testbed at the Space Telescope Science Institute. HiCAT does not have an off-axis light source to emulate a planet on the testbed, so planets are injected via software into the raw testbed images. The AI models were trained on the incoherent estimate images resulting from the wavefront sensing and control to distinguish planetary signals from noise, such as starlight leakage, and to identify potential planets within an image. These AI models also produce per-pixel confidence maps indicating the likelihood of a planetary signal. Once the AI models identify a possible planetary signal, the cell mapping and all appropriate EKF components are adjusted to center the potential planet within a single cell. This enables the EKF to refine the estimation of coherent and incoherent light at the updated grid coordinates, supporting a more dynamic adaptation to changing wavefront errors and helping to ensure optimal suppression of residual starlight. This work introduces a closed-loop architecture in which AI-driven detections are used to adapt the EKF state-space representation in real time, re-centering the cell mapping around suspected planetary locations. This enables the wavefront control system to locally refine the dark zone while preserving sensitivity to faint incoherent sources. This CNN-EKF and ViTEKF approach provides a pathway for integrating machine learning and statistical modeling for real-time, autonomous exoplanet detection. It highlights the potential for advancing the performance of future space-based high-contrast imaging missions.
MIT Department
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics
Terms of Use
Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)
Copyright retained by author(s)
Persistent DSpace Link
