Line of research overview

Accelerating the dark matter signal search in astrophysical data with machine learning

Principal Investigator: Gabrijela Zaharijas

Abstract: Over the past fifteen years, several satellites exploring the high-energy (HE) sky with gamma rays and charged cosmic rays gathered unprecedented data on the most energetic astrophysical processes in our Galaxy. These measurements resulted in a series of exciting discoveries in the field of HE astrophysics and provided an opportunity to search for the nature of Dark Matter (DM) particles in a completely new way. One of the limitations to taking the full advantage of this high-quality data is related to the challenges of thorough explorations of multi-dimensional parameter spaces. At the same time, machine-learning (ML) algorithms are being increasingly used to tackle the analyses of large data sets. They are now evolving to include elements of statistical analysis, which makes them suitable to address the science questions, as the one above and could bring further breakthroughs in the field. Inspired in part by the arguments listed above, the scientific community joined together within the framework of darkmachines.org, which connects data scientists with physicists and astrophysicists, with a motivation to ‘accelerate search for dark matter with machine learning’. The community already had two face-to-face meetings, one in Leiden in 2018, and one organized also in part by IFPU, in Trieste in 2019. 

This program focuses on an attempt to improve the analysis of the astrophysical data by adopting the ML techniques, with the final motivation to discern signals of particle DM. Some of the projects in this research program, currently under development or planned in the short time scale include: – localization and classification of the gamma-ray point sources using a combination of supervised and unsupervised learning with the goal of looking for new source classes, including possibly the one from DM subhalos  – analysis of the stellar or strong lensing data with the goal to constrain small scale DM clustering, – search for new physics with gravitational waves using ML. We anticipate that this line of research will likely expand to include more applications in the future.

Status of project and perspectives: Over the past few years the group developed AutoSourceID (ASID) ML suite that localizes and classifies point sources from raw astrophysical images. It was applied to gamma ray (Fermi LAT) [2] and optical (Meer-Licht) [1] data and proved to outperform traditional algorithms in a range of aspects. We’ve also shown that networks trained on one wavelength perform well in a range of different wavelength bands, and as the next step, we plan to train networks that use multi-wavelength information. We also use DeepEnsamble networks to determine properties of large scale gamma ray emission, focusing in particular on the region close to the Galactic center.

In the future we plan to use these algorithms to look for structure in optical data to search for signatures of DM clustering on Galactic scales and to explore their potential on searches for new physics in gravitational wave data.