PhD Defense: Loris Berthelot

Titre: “Deep learning for Galactic interstellar filament detection.”

jury:

Stella OFFNER, COSMICAI, Université du Texas à Austin — Rapporteure
Patrick GALLINARI, ISIR, Sorbonne Université — Rapporteur
Cécile CAPPONI, LIS, Aix-Marseille-Université — Présidente
Alexis LECHERVY, GREYC, Université de Caen Normandie — Examinateur
Marc HUERTAS-COMPANY, IAC, Université de La Laguna — Examinateur
Thierry ARTIÈRES, École Centrale Méditerranée/LIS, Aix-Marseille-Université — Directeur de thèse
Annie ZAVAGNO, LAM, Aix-Marseille-Université — Directrice de thèse
François-Xavier DUPÉ, LIS, Aix-Marseille-Université — Co-encadrant de thèse
Doris ARZOUMANIAN, AIS, Kyushu University — Membre invitée
Eugenio SCHISANO, IAPS, Istituto Nazionale di Astrofisica — Membre invité

Summary:

In galaxies, stars form within filaments made of gas (primarily hydrogen) and dust (small solid particles composed mainly of carbon). These filaments arise in the interstellar medium (the space between stars) and evolve according to the surrounding physical conditions and the star formation they host. They are observed with ground- and space-based instruments and studied through numerical simulations. Detecting these structures in data—both observational and simulated—is a necessary prerequisite for investigating the very earliest phases of star formation. Many algorithms have been developed to detect such filaments. They rely on parameters that describe the properties of the target filaments, in particular their contrast with respect to the surrounding emission. Results show that it is difficult to detect all filaments. In particular, low-contrast and/or low-density filaments are often missed, leading to a detection bias toward brighter filaments or toward fainter filaments that happen to lie in regions of low background emission. Moreover, these methods are sensitive to parameter settings that depend strongly on the data. In light of this, exploring the potential of deep learning has been proposed—this is the context in which the present inter-disciplinary PhD was carried out. The first part of this work presents contributions focused on training and applying neural networks to a very large image of the Galactic Plane of the Milky Way for a semantic segmentation task. We implemented a spatial tiling strategy combined with semi-supervised learning to address astrophysical data specificities (data volume, large variations in intensity and contrast, and incomplete ground truth). In addition, we proposed a new U-Net variant, called PE-UNet, which explicitly incorporates the galactic position of each example as auxiliary information during training. Experiments show that PE-UNet significantly improves segmentation performance over the studied architectures, conclusions that are supported by an in-depth astrophysical analysis of the segmentation maps produced by the different models.
Because current algorithms fail to detect faint filaments—and their outputs are used as training labels for neural networks—it becomes impossible to use these undetected filaments for either training or evaluation, which strongly biases the results. For this reason, the second part of the thesis relies on the physical modeling of filaments using the Plummer radial profile and on generating a synthetic dataset based on this model to counter the biases present in observational data. This synthetic dataset enables the evaluation of a wide range of filament-detection methods—from classical algorithms to semantic and instance segmentation models. It also serves to analyze several learning biases typically encountered when models are trained on biased observations. Finally, the dataset makes it possible to design a processing pipeline that estimates the likelihood that a given prediction truly corresponds to a filament by leveraging prior physical knowledge. This pipeline explores Physics-Informed Neural Networks (PINNs) with the goal of injecting physical knowledge into neural networks.

Keywords: Deep learning – Semantic segmentation – Stellar Formation