Revealing Galactic Structure with Simulation Based Inference

Abstract

The structure of the Milky Way is poorly constrained. There is strong evidence that the Milky Way has spiral features, but the location and morphology of these features are largely unknown and rarely constrained quantitatively. We develop a simplistic probabilistic model of a multi-armed, log-spiral Galaxy that predicts the Galactic longitude, latitude, and LSR velocities of massive star forming regions associated with spiral arms. The likelihood distribution for this model is intractable since it is marginalized over the unknown Galactocentric azimuths (i.e., distances) of the nebulae. We therefore train a neural network to predict the likelihood distribution through a machine learning technique known as simulation based inference (SBI). With the learned likelihood distribution, we infer the model parameters using Monte Carlo Markov Chain (MCMC) methods. We demonstrate the accuracy of the learned likelihood distribution and the success of our SBI+MCMC procedure in recovering the “true" model parameters for a synthetic, four-armed, log-spiral Galaxy. Preliminary attempts to infer the model parameters from the latest Galactic HII region data reveal model degeneracies. SBI is a viable tool to test sophisticated models of Galactic structure on position-velocity data.