Data-driven modeling of rotation curves with artificial neural networks

TL;DR

The paper tackles the challenge of deriving galaxy mass distributions from rotation curves by introducing per-galaxy data-driven ANNs that predict $V(r)$ and its uncertainty without imposing predefined mass profiles. The approach leverages Monte Carlo Dropout for uncertainty quantification and Genetic Algorithms to tailor network architectures to each galaxy, training on THINGS RC data. Comparisons with parametric NFW models under fixed and free $\Upsilon_{\star}$ reveal three groups of galaxies: those well-described by parametric profiles, those where ANN offers intermediate improvements, and those (Group C) where ANN substantially outperforms both parametric fits, especially for complex or irregular mass distributions. Overall, the study demonstrates that data-driven models can complement traditional theories, identify cases where standard descriptions fail, and guide future refinements of baryon–dark matter modeling in spiral galaxies.

Abstract

Galactic rotation curves are crucial for understanding the distribution of mass in galaxies. Despite advances in precision observations, there are discrepancies between the inferred mass from luminosity and the observed rotational velocities, often attributed to dark matter. While traditional parametric models provide valuable insights, they struggle with complex galactic features like prominent bulges and non-circular motions. In this study, we apply artificial neural networks to generate robust, data-driven models, tailored to each galaxy, for the rotation curves of spiral galaxies using high-quality observational data. Our approach demonstrates that neural networks can effectively capture the intricate structure of rotation curves without relying on predefined astrophysical assumptions. By comparing the data-based models with the Navarro-Frenk-White model under two different assumptions for the stellar component, we classify galaxies based on the model that best fits their rotation curves, offering insights into the limitations and strengths of both theoretical and data-based methods. This work highlights the potential of machine learning techniques in identifying galaxies whose dynamics are not well captured by standard theoretical models, pointing to the need for more refined physical descriptions.

Figures (3)

• Figure 1: ANNs general architecture for the rotation curves. The number of hidden layers and their number of neurons depend on each galaxy because for each case a different neural network model is generated. In all the cases, the input is the values of the radius, and the output is the velocity and the statistical error.
• Figure 2: Neural network models for the 17 galaxies in comparison with the observational data and the NFW profiles $\Upsilon_{\star}^{free}$ and $\Upsilon_{\star}^{fix}$. The rotation velocity (y-axis) is given in km/s, whereas the radius (x-axis) is given in Kpc.
• Figure 3: Parametric mass contributions inferred from observed (Obs) and data-driven (ANN) rotation curves for three representative galaxies of groups A, B, and C. For each system, the upper panel shows the halo and stellar contributions obtained from independent NFW fits with free stellar mass-to-light ratio, while the lower panel displays the differences between the ANN and observation based components, $\Delta V(r) \equiv V_{\rm ANN}(r) - V_{\rm obs}(r)$. Blue curves correspond to the dark matter halo, red curves to the stellar components, while solid and dashed lines indicate fits to the observed and data-driven rotation curves, respectively.