Learning from galactic rotation curves: a neural network approach

TL;DR

This paper develops neural network models trained on simulated galaxy rotation curves to accurately infer dark matter and baryonic parameters from observed data, accounting for observational noise and uncertainties.

Contribution

It introduces a neural network approach for direct parameter inference from galaxy rotation curves, incorporating noise effects and uncertainty quantification, advancing beyond traditional methods.

Findings

Neural networks accurately infer dark matter and baryonic parameters from observed rotation curves.

Training on simulated data with noise improves robustness of parameter estimation.

Uncertainty quantification methods compare favorably with Bayesian approaches.

Abstract

For a galaxy, given its observed rotation curve, can one directly infer parameters of the dark matter density profile (such as dark matter particle mass $m$, scaling parameter $s$, core-to-envelope transition radius $r_t$ and NFW scale radius $r_s$), along with Baryonic parameters (such as the stellar mass-to-light ratio $\Upsilon_*$)? In this work, using simulated rotation curves, we train neural networks, which can then be fed observed rotation curves of dark matter dominated dwarf galaxies from the SPARC catalog, to infer parameter values and their uncertainties. Since observed rotation curves have errors, we also explore the very important effect of noise in the training data on the inference. We employ two different methods to quantify uncertainties in the estimated parameters, and compare the results with those obtained using Bayesian methods. We find that the trained neural networks can extract parameters that describe observations well for the galaxies we studied.