The formation of stars is a fundamental aspect of the evolution of the Universe. Strangely it is still poorly understood as it results from an intricate combination of complex physical processes : several instabilities (dynamical, chemical, thermal), magneto-hydrodynamic turbulence, gravity and energy injection by stars themselves. Because of this complex, multi-scale and multi-phase physics, this problem is now study using numerical simulations of increasing sophistication. Important progress has been made to a point that we are now in a situation where there is a lack of constraints from observations to identify the relevant physical scenarios. This is in part due to the difficulty to find the right metric to compare numerical simulations and observations. With upcoming massive data sets, in particular hyper-spectral data obtained with radio-telescopes like ALMA and soon SKA, new opportunities and challenges arise.
In this context, machine learning tools open new ways of exploration. In particular they allow the automatic definition of new comparison and evaluation metrics obtained by training on a large number of systems that share some physical properties. It now becomes possible to quantify the statistical properties of various physical situations and identify physical properties of the interstellar medium by comparing directly with a set of simulations.
The subject of this Ph.D. thesis is to define a framework that would allow to estimate physical parameters (magnetic field intensity, temperature distribution, density distribution, power spectra of density and velocity) of different regions of the interstellar medium by applying machine learning tools on hyper-spectral observations (21 cm and CO). The setup of the tools would be first done by a training on a set of numerical simulations that are already available. This project is made possible thanks to a combination of expertise present in the Hyperstars collaboration , a joint effort by experts of the star formation process (Marc-Antoine Miville-Deschênes - hyper-spectral data and Patrick Hennebelle - numerical simulations) and experts of several aspects of data science, from machine learning to GPU implementation.
Application should be done through ADUM
For more information, please contact Marc-Antoine Miville-Deschênes (mamd at cea.fr).