Reference : UMR7326-ANAMEK-047
Workplace : MARSEILLE 13
Date of publication : Thursday, September 15, 2022
Scientific Responsible name : ZAVAGNO Annie (LAM) – ARTIÈRES Thierry (LIS)
Type of Contract : PhD Student contract / Thesis offer
Contract Period : 36 months
Start date of the thesis : 1 November 2022
Proportion of work : Full time
Remuneration : 2 135,00 € gross monthly
Description of the thesis topic
Title : Deep learning for the study of galactic star formation
Star formation is a multiscale process that controls the evolution of galaxies. In our Galaxy, available data reveal this process on spatial and temporal scales spanning more than 6 orders of magnitude. This large mass of data, combined with information available on the physical conditions of the medium (density, temperature, turbulence, magnetic field) and numerical simulations, offer the opportunity to establish the link between the properties of the stellar formation and those of the medium which hosts it. More specifically, star formation in galaxies takes place in filamentary structures composed of gas (mainly hydrogen) and dust (small carbonaceous solid particles). These filaments form in the interstellar medium (medium located between stars) and evolve until they fragment to form pre-stellar cores that will house future stars. Although these filaments have been very widely studied, their formation and evolution are still very poorly understood, in part because a comprehensive study of all their evolutionary stages has not yet been carried out. Such a study requires combining the knowledge available on these filaments (data and models). However, the volume and complexity of the data available (in particular the arrangement of multiple views of the data) make this combination inaccessible to “classic” human research. Current approaches, whether based on mathematical morphology, image processing, neural networks, do not integrate physical knowledge and are developed or learned from data alone. However, physical models exist and can be integrated in different ways in the construction of a filament detection method, whether for the construction of artificial data sets by simulation, by the direct integration of knowledge into a neural architecture (a major challenge today) or in the definition of an adequate cost function to favor inferences more faithful to reality (via adversarial strategies for example). The objective of this PhD Thesis is to develop innovative machine learning methods for the detection of filaments. One of the first steps will aim to assess the effectiveness of state-of-the-art methods on large masses of data with a low number of views (two or three wavelengths). The avenues envisaged to go beyond the state of the art consist in particular in the hybridization of deep neural networks and mathematical morphology operators or the development of original neural models operating on graphs. To go further we will study sequential acquisition methods to allow a scaling up of the number of views used to make a decision. Reinforcement algorithms or end-to-end optimizable formalizations can be considered for this. These approaches have the particular advantage of being able to adapt to the case of missing views. The doctoral student will focus on the methodological aspects of machine learning developed for the astrophysical data studied. The preparation of the datasets and the astrophysical knowledge will be provided by the various actors of a larger project in which this PhD Thesis is funded. The student will of course be closely associated with these developments.
Available funding for computer and missions.
International collaboration with Italy (Rome, Naples).
Important: Recognized experience (project, internship) in Machine Learning / Deep Learning and/or Astrophysics is required to apply for this position. Applications with no experience in any of these areas will not be considered.
We talk about it on Twitter!