For the first time, two young researchers from the Strasbourg Astronomical Observatory have created an artificial intelligence to estimate the properties of globular clusters observed in galaxies, including the Milky Way.
Globular clusters are collections of tens of thousands to several million stars in the form of compact spheres bound together by the force of gravity. These star clusters are among the oldest objects in the Universe and have survived 13 billion years of evolution. This is why the study of these objects is fundamental for the understanding of the formation and history of galaxies. Obtaining an accurate estimate of their properties will allow the scientific community to build more and more precise scenarios about the past of our Milky Way and other distant galaxies.
About 160 globular clusters are observed in our Galaxy but it is very difficult to estimate their properties in detail. The problem is that these systems are very old and have therefore undergone a long dynamical evolution as well as a long stellar evolution. For this reason, several stars within the clusters have had time to escape or reach the end of their life, for example by transforming into black holes, which makes them undetectable to observation. Consequently, obtaining good constraints on the properties of clusters (for example, their mass, age or distance from Earth) is complicated: to date, this requires the development of very elaborate dynamic models and requires a significant amount of computing time.
To overcome this problem, the two researchers from Strasbourg have developed the “π-DOC” (Predicting Images for the Dynamics Of stellar Clusters) algorithm. In an article to be published in the “Monthly Notices of the Royal Astronomical Society”, they have shown that convolutional neural networks allow to retrieve the distance, age and mass of these objects, using the image of the clusters (their “luminosity map”) as the only input. To train their algorithm they used numerical simulations of these objects to generate “mock” observations of their luminosity maps as they would be observed by a real telescope. Once trained, π-DOC predicts, almost instantaneously, the properties of the clusters with a low error, comparable to the best models currently available (Figure 1).
Even better, a first application of π-DOC to observations has shown that the algorithm already gives good estimates on real astronomical data (Figure 2).
This considerable work is made possible thanks to one of the largest supercomputers in France, Jean-Zay (GENCI), on which the two researchers developed their project. Ultimately, their aim is to train their algorithm on a more extensive set of simulations in order to incorporate more examples that resemble reality. In this way, the algorithm should find very quickly and with unequalled precision the properties of all the clusters already known and it will make it possible to systematically obtain the “identity card” of the new clusters observed by future generations of telescopes.