A simulation-based inference framework for Milky Way satellite analyses

The Vera C. Rubin Observatory has just begun a decade-long survey of the southern sky that is expected to discover many of the faint dwarf galaxies orbiting our own Milky Way. These tiny neighboring galaxies are among the most dark-matter-dominated objects known, making them powerful natural laboratories for studying dark matter — the invisible substance that makes up most of the matter in the Universe and whose true nature remains one of the biggest open questions in science. This project will develop new computational tools, powered by recent advances in artificial intelligence, that let researchers compare theoretical predictions for these dwarf galaxies against real observations with unprecedented rigor. Using them, the team from the University of New Mexico will test whether dark matter behaves as the leading “cold dark matter” theory predicts, or whether it has hidden properties that reveal themselves only on the smallest cosmic scales — a finding that would reshape decades of thinking about the Universe. The project supports the training of a graduate student and, through a mentoring program at the University of New Mexico, brings cutting-edge AI and machine-learning skills to early-career undergraduates, helping prepare a proficient STEM workforce. This program advances quantitative, model-independent tests of the two pillars of the cold dark matter (CDM) paradigm — that dark matter is cold (clustering on all scales) and collisionless (interacting only gravitationally) — using the Milky Way satellite population that the Rubin Observatory’s Legacy Survey of Space and Time will greatly expand. The work has two objectives. First, the team will extend its fast semi-analytic forward model of satellite luminosities, sizes, and stellar velocity dispersions to self-interacting dark matter (SIDM), capturing the orbital, tidal, and gravothermal evolution of subhalos via machine-learning emulators trained on semi-analytic codes and cross-checked against N-body simulation suites. This model will drive a likelihood analysis of currently known satellites and forecasts of the constraining power of post-survey spectroscopic follow-up. Second, to overcome the intractable likelihoods and large latent space of the satellite population, the team will develop a simulation-based inference framework coupling graph neural networks to normalizing flows, learning the joint posterior of dark matter and galaxy-formation parameters directly from data and fully mapping their degeneracies. Together these tools will sharpen the reach of Milky Way satellites as astrophysical probes of dark matter microphysics. The project supports a PhD student and the development of graph-neural-network, normalizing-flow, and simulation-based-inference teaching modules for UNM’s PRIMER undergraduate mentoring program. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.