Actuality from the project

Detection and Characterization of Exoplanets via Direct Imaging:
We have developed new approaches for analyzing high-contrast imaging observations. By combining data-driven statistical learning methods with a signal processing framework, our methods are interpretable, and the characterization of uncertainties is better controlled. Previous work was limited to modeling a single observation sequence. We applied a deep learning method to a large archive of observations to build a more accurate and robust model of direct imaging data, leading to increased sensitivity, particularly at small angular separations from the host star.

Reconstruction of circumstellar disks:
Direct imaging allows us to reconstruct an image of a star’s immediate surroundings, thereby enabling us to image protoplanetary disks in young systems. We have developed algorithms that extract information using spectral and angular diversity (the ASDI technique), achieving unparalleled reconstruction quality. We have demonstrated that it is possible to separate the contributions of a circumstellar disk and exoplanets, thereby providing astronomers with much more precise information on the interaction between the disk and exoplanets.

Processing Multi-Epoch Observations:
The detection limits for exoplanets can be pushed further by combining observations taken several months or years apart. The motion of exoplanets along their orbits provides essential information about their mass and allows us to rule out false detections by accepting only trajectories that are consistent with Kepler’s laws.

Combining Instrument Models and Disk Physics:
The reconstruction of circumstellar disks from polarimetric measurements taken by the VLT’s SPHERE/ZIMPOL instrument was achieved by combining a detailed consideration of the instrument’s optics with expertise in disk modeling. Simulators enable the generation of a very large number of disk images based on physical principles. An AI model was trained on these images to capture the statistical relationships within them and was then integrated into a reconstruction algorithm. The reconstructed disks are thus more physically realistic, and the proposed approach allows for the characterization of prediction uncertainties.

Efficient Learning of Dense and Non-Stationary Covariances:
Modeling correlations is central to many applications. In very high-dimensional settings, this task is challenging because the number of parameters becomes enormous. It is therefore essential to impose a structure on the covariance. We have developed models that are more general than classical models (diagonal, circular, or block-diagonal) and highly efficient for both parameter estimation and model application. In high-contrast imaging, these new models make it possible to push the limits of sensitivity in exoplanet detection. In conjunction with the “Unsupervised AO Control” project under the ORIGINS PEPR, we will also explore the application of these covariance models for adaptive optics control.

Probabilistic Models for Image Reconstruction Subject to Positivity Constraints:
We have developed statistical inference methods tailored to high-dimensional signal processing that model the positivity of astronomical images. These models are designed to yield tractable algorithms, particularly for image deconvolution applications.

Co-design of high-angular-resolution systems assisted by adaptive optics:
When a high-angular-resolution imaging system combines adaptive optics and deconvolution algorithms to minimize blur as much as possible, optimal performance can only be achieved through a global (i.e., joint) optimization of the entire imaging chain. We have developed approaches for selecting specific system parameters to achieve this optimal configuration, demonstrating a significant improvement over separate optimization of the adaptive optics system and the image restoration algorithm.

Focus

Detection and Characterization of Exoplanets Using the Radial Velocity Method:
The analysis of spectroscopic measurements requires solving a difficult inverse problem: it is nonlinear, of unknown order (the number of exoplanets must be determined from the data), and subject to various perturbations (stellar activity, terrestrial effects, etc.). We plan to implement advanced signal processing methods (Bernoulli-Gaussian stochastic models) to improve the analysis of these data.

Statistical Learning from Large Archives of Observations for the Reconstruction of Circumstellar Disks:
In high-contrast imaging, it is very difficult to separate the signal of interest (the image of the disk) from the stray light coming from the host star. This task can be learned from the numerous available archive images.

Self-Supervised Learning for Image Reconstruction in Astronomy:
Most AI methods rely on training datasets for which the expected response is known (supervised learning). In scientific imaging, it is essential to be able to train models without introducing bias related to expected responses—that is, to learn solely from observations. Self-supervised learning is a very active area in statistical learning and computer vision that warrants integration with the specific characteristics of astronomical observation to yield methods applicable to real-world data.

Merging of High Angular Resolution and High-Contrast Data:
High-contrast imaging (behind a coronagraph) and interferometry (Sparse Aperture Masking, long-baseline interferometry) provide access to complementary regions for observing complex circumstellar disks: the outer region via high-contrast imaging, and the inner region accessible only through interferometry. The fusion of these two complementary modalities will provide access to information that is crucial for understanding the mechanisms at work within these disks.

Relation with others WP

The statistical models developed are also applied to the problem of predictive control of adaptive optics systems (the “Unsupervised AO Control” project under the PEPR ORIGINS program).

Relation with industries

not applicable at this time

Publication links to the project

[1]               Romain Fétick, Alix Yan, Laurent M. Mugnier, Jean–François Giovannelli, Cyril Petit. Robust myopic deconvolution combining the parametric marginal approach and a support constraint: experimental validation on Adaptive–Optics corrected images. AOE4LT8, Oct 2025, Vina del Mar, Chile. 2025. ⟨hal–05542387⟩

[2]               Florian Cheyssial, Laurent M. Mugnier, Cyril Petit. Stratégie de co–conception pour les systèmes d’optique adaptative : application à l’optimisation des longueurs d’onde pour l’observation de satellites. GRETSI 2025, Aug 2025, Strasbourg, France. pp. 1141–1144 / 2025–1747. ⟨hal–05521246⟩

[3]               Pierre Minier, Jean–François Giovannelli, François Orieux. A Probabilistic Model for Image Processing with Positivity Constraint and Spectral Density Control. 2025 IEEE Statistical Signal Processing Workshop (SSP2025), Jun 2025, Edimbourg, United Kingdom. ⟨10.1109/ssp64130.2025.11073477⟩. ⟨hal–05049694⟩

[4]               Pierre Minier, Jean–François Giovannelli, François Orieux, Marcelo Pereyra. Sampling High–Dimensional Constrained Gaussian Distributions Using Circulant Gibbs. IEEE International Conference on Image Processing (ICIP 2026), Sep 2026, Tempere, Finland. pp.1–6, ⟨10.1109/ICIP61757.2026.11630488⟩. ⟨hal–05644009⟩

[5]               Olivier Flasseur, Loïc Denis, Éric Thiébaut, Maud Langlois. REXPACO ASDI: Joint unmixing and deconvolution of the circumstellar environment by angular and spectral differential imaging. Monthly Notices of the Royal Astronomical Society, 2024, 535 (1), pp.689–728. ⟨10.1093/mnras/stae2291⟩. ⟨hal–04719549⟩

[6]               Yann Gutierrez, Johan Mazoyer, Laurent M. Mugnier, Olivier Herscovici–Schiller, Baptiste Abeloos. Image–based wavefront correction using model–free Reinforcement Learning. Optics Express, 2024, 32 (18), pp.31247. ⟨10.1364/OE.529415⟩. ⟨hal–04624895v2⟩

[7]               Quentin Villegas, Laurence Denneulin, Simon Prunet, André Ferrari, Nelly Pustelnik, et al.. Modèles de diffusion pour la reconstruction polarimétrique d’environnements circumstellaires. GRETSI 2025 – XXXe Colloque sur le Traitement du Signal et des Images, Aug 2025, Strasboug, France. ⟨hal–05308782v2⟩

[8]               Théo Bodrito, Olivier Flasseur, Julien Mairal, Jean Ponce, Maud Langlois, et al.. A New Statistical Model of Star Speckles for Learning to Detect and Characterize Exoplanets in Direct Imaging Observations. CVPR 2025 – IEEE / CVF Conference on Computer Vision and Pattern Recognition, Jun 2025, Nashville, United States. pp.1–15. ⟨hal–05000360⟩

[9]               Théo Bodrito, Olivier Flasseur, Julien Mairal, Jean Ponce, Maud Langlois, et al.. Deep learning for exoplanet detection and characterization by direct imaging at high contrast. JDLS 2025 : Troisième Journée Deep Learning For Science, Jun 2025, Paris, France. pp.1–1, 2025. ⟨hal–05251768⟩

[10]             Théo Bodrito, Olivier Flasseur, Julien Mairal, Jean Ponce, Maud Langlois, et al.. Deep learning for exoplanet detection and characterization by direct imaging at high contrast. SF2A 2025 – Journées de la Société Française d’Astronomie & d’Astrophysique, Jul 2025, Toulouse, France. pp.1–5. ⟨hal–05281730⟩

[11]             Théo Bodrito, Olivier Flasseur, Julien Mairal, Jean Ponce, Maud Langlois, et al.. Modèle statistique apprenable de mélange de distributions et fusion de données multivariées pour l’imagerie d’exoplanètes. GRETSI 2025 – XXXe Colloque Francophone de Traitement du Signal et des Images, Aug 2025, Strasbourg, France. pp.1–4. ⟨hal–05140005⟩

[12]             Théo Bodrito, Olivier Flasseur, Julien Mairal, Jean Ponce, Maud Langlois, et al.. Joint statistical modeling and deep learning for exoplanet detection and characterization by direct imaging at high contrast. EPSC–DPS Joint Meeting 2025, Sep 2025, Helsinki, Finland. pp.1–47. ⟨hal–05251737⟩