Colloquium by Lina Necib, MIT
In this talk, I will explore the interfacing of simulations, observations, and machine learning techniques to construct a detailed map of dark matter in the Milky Way, and argue that progress now hinges on the dance between the information content of our data and the fidelity of our models. For the Galactic halo, I will present recent work leveraging Gaia DR3 that reveals a declining stellar rotation curve. I will then show how underestimated errors can bias the resulting curve, and how we used simulations to build a more robust error treatment that reconciles the measurement with other probes of the Galaxy's enclosed mass. In dwarf galaxies, stellar data of the faintest systems will not substantially grow and therefore I present GraphNPE, a graph neural network methodology that accurately extracts dark matter density profiles from sparse stellar kinematics. With GraphNPE, we find a hint of a core in Boötes I, consistent with predictions of self-interacting dark matter. Finally, I will invert the question and quantify, using mutual information, how the memory of the Milky Way's merger history is progressively erased from stellar dynamics, setting a fundamental limit on what Galactic archaeology can recover. Together, these results outline a program of inference that seeks to extract the most robust model of dark matter in the Milky Way from available data, arriving just as Gaia DR4 and the next generation of surveys come online.
Cookies will be available, starting at 3:45.
Host: Kathryn Johnston