Low Galaxy Radii Dark Matter Predictions

A Hybrid Attention based Learning and Classical Modeling approach.

Abstract

The discrepancy between observed galaxy rotation curves and calculations via Newtonian
dynamics has long evidenced the presence of dark matter. Classically represented dark
matter halo models, such as the Navarro-Frenk-White (NFW), Einasto, and DC14 profiles,
are frequently used to infer the distribution of this unseen mass. Whilst effective at large
galaxy radii, these models often fail to accurately represent the shape of rotation curves
at central regions of a galaxy. This failure arises from fixed parametric forms within the
models that cannot account for the complex interplay between baryonic physics and dark
matter.
This literature review explores the embedding of transformer-based machine learning
(ML) models within classical models to improve the representation of dark matter inferences.
Specifically, transformers trained on galaxy feature datasets learn to predict the parameters
retained within these analytic profiles; optimising their fit via learnable weights. By
introducing self-attention mechanisms, transformer models identify contextual inter
dependencies between observational inputs—such as HI linewidths, stellar densities, and
scale radii—and output refined parameter estimates that resolve low-radius modeling
errors