Inferring the Structure of Galaxies with Deep Convolutional Neural Networks

The popular and beautiful galaxy images provided from telescopes like the NASA’s Hubble Space Telescope or the Sloan Digital Sky Survey, enabled astronomers to a plethora of studies that greatly improved our knowledge of the Universe and its evolution. In particular the structure of the galaxies can be effectively characterized from their surface light profile distribution through a technique called profile fitting, consisting in using analytic function to obtain a set of simple parameters that would ideally allow the reconstruction of the 2D photometrical shape of the galaxy. Having the parameters decomposition of large data-samples of galaxies at different cosmic ages, allow to understand how galaxies evolve and how they interact with the gas, dust and the other galaxies in the surrounding. These studies are closely related with question on the origin and evolution of our Universe.

140110105701-faraway-galaxies-horizontal-large-gallery.jpeg

Modern and future large area surveys will provide statistics for billions of galaxies and astronomical sources, like never have been possible in the past. This data have the potentiality to revolution Astronomy and our knowledge of the Universe, however classical algorithms used by astronomers to model the light profiles of galaxies, have not been conceived to deal with such large amount of data, and they are totally unsuited for the amount of data that will be provided.

GAMOCLASS project: In this project we applied convolutional neural networks to develop a fast and reliable method for galaxy profile fitting. In scientific literature, deep neural network methods have never been applied to this kind of problems. In contrast with traditional methods used in astronomy, our approach does not require any tuning previous to the application of the algorithm but only needs realistic simulations to train the neural network. Our code provides results having comparable accuracy than traditional methods, but, once trained, our code is in the order of 400 times faster. The drastic cut down of the required computational time, render our code ideal to deal with very large astronomical dataset.

stamp_3images.jpg

Team and funding: The project is funded by PSL-Research University and involves researchers of the Observatoire de Paris and the Centre for Mathematical Morphology, MINES ParisTech. The team of researchers involved in this project is composed by Etienne Decencière, Marc Huertas-Company, Diego Tuccillo and Santiago Velasco-Forero.

Reference:

Tuccillo, D., Huertas-Company, M., Decencière, E., Velasco-Forero, S., Domínguez Sánchez, H., & Dimauro, P. (2017). Deep learning for galaxy surface brightness profile fitting. Monthly Notices of the Royal Astronomical Society, 475(1), 894-909.

%d bloggers like this: