Beyond single tracers: CNN-based inference of galaxy mass profiles from combined gas and stellar kinematics

TL;DR

This work addresses the challenge of inferring galaxy mass profiles by leveraging both gas and stellar kinematics. It introduces a probabilistic convolutional neural network with a normalizing flow that processes star- and HI-based input maps, trained on realistic cosmological hydrodynamical simulations (NIHAO and AURIGA) and evaluated across multiple radii. The key finding is that combining gas and stellar tracers reduces the scatter in mass predictions by up to about $\sim 1.5$ relative to single-tracer inputs, though generalization across simulation suites remains nontrivial and benefits from mixed-suite training. The approach offers a data-driven, uncertainty-aware pathway toward dynamical mass inferences that can be extended to observational data, with careful attention to simulation-dependent systematics.

Abstract

We investigate whether combining gas and stellar kinematic maps provides measurable advantages in recovering galaxy mass profiles, compared to using single-component maps alone. While traditional methods struggle to integrate multi-tracer data effectively, we test whether deep learning models can leverage this joint information. We develop a probabilistic convolutional neural network (CNN) framework trained and tested on mock galaxy kinematic maps from multiple cosmological simulation suites. Our model is trained on gas-only, stars-only, and combined gas+stellar velocity maps, allowing direct comparison of performance across tracers. To assess robustness, we include simulations with differing feedback models and galaxy properties. Combining gas and stellar maps reduces the dispersion in the inferred mass profiles by up to a factor of $\sim$1.5 compared to models using either tracer independently. The CNN architecture effectively captures complementary information from the two components. However, we find limitations in generalizing between simulation suites, with reduced performance when applying models trained on one suite to galaxies from another.

Figures (9)

• Figure 1: Star information input maps for a single edge-on galaxy from the NIHAO sample of M$_* = 10.68$ M$_\odot$. Top: PDF in the {$\hat{R}_{\rm proj},\hat{v}_{\rm LOS}$} phase space. Bottom: PDF in the {x,y} phase space. Both obtained following the procedure described in section \ref{['sec:network:input:star']}.
• Figure 2: Gas information input maps for a single galaxy from the NIHAO sample of M$_* = 10.68$ M$_\odot$ at increasing inclinations from edge-on to 60 degrees. From left to right: HI intensity, average line-of-sight velocity and velocity dispersion maps, obtained following the procedure described in section \ref{['sec:network:input:gas']}.
• Figure 3: Scheme of our CNN architecture. Five parallel branches (two for star information, three for gas information) pass through several convolutional and pooling layers independly, further reducing the dimensionality, and then they merge for a final sequence of dense and pooling layers, producing a N parameter output. After training the CNN, the 32 neurons of penultimate layer are used as inputs to train a normalising flow model. The flow model learns a series of transformations to a N-dimensional gaussian PDF, which are conditioned on the inputs, and outputs a posterior N-dimensional joint PDF. The final output represents the value estimated for the dynamical mass of the galaxy enclosed within N different radii.
• Figure 4: Ratio between the mass predicted by the neural network models and the real mass enclosed within different radii of the galaxies in the test sets of NIHAO galaxy projections. Top: Results for the Gas model. Bottom: Results for the StarGas model.
• Figure 5: Ratio between the mass predicted by the neural network models and the real mass enclosed within different radii of the galaxies in the test sets of NIHAO galaxy projections. The colored lines show the median ratio using the mass estimated by the CNN for the training set, while the shadowed regions indicate the 1-$\sigma$. Blue: Results when using the three gas channels described in Sec. \ref{['sec:network:architecture']}, the radius where the cold gas superficial density falls below 1 M$_\odot$ pc$^{-2}$ and the mean velocity dispersion inside that radius. Green: Results when using the two star channels described in Sec. \ref{['sec:network:architecture']}, the projected half-light radius and the mean velocity dispersion of the stars. Red: Results when using all of the above. The predicted to true enclosed mass ratios resulting from applying literature mass estimators to NIHAO galaxies are shown as black symbols with 1-$\sigma$ errorbars.