On the sensitivity of different galaxy properties to warm dark matter

TL;DR

This paper investigates how sensitively galaxy properties respond to warm dark matter by analyzing 1,024 DREAMS cosmological hydrodynamical simulations. It uses a combination of machine learning methods—an MLP with normalizing flows on global galaxy-property statistics and a GNN with normalizing flows on halo substructure—to infer the WDM mass, complemented by symbolic regression for interpretable relations. The key finding is that the gas content of subhalos, encapsulated in the gas-mass distribution and its zeros, provides the strongest constraint on $m_{\rm WDM}$ (with $R^2$ up to ~0.94 when using global statistics), while halo-level information adds only marginal improvements beyond these global descriptors. The results suggest that large-scale galaxy-population statistics largely govern WDM sensitivity, with potential extensions to observational probes such as HI, and they demonstrate robust methodology for constraining dark matter physics from complex galaxy formation simulations.

Abstract

We study the impact of warm dark matter (WDM) particle mass on galaxy properties using 1,024 state-of-the-art cosmological hydrodynamical simulations from the DREAMS project. We begin by using a Multilayer Perceptron (MLP) coupled with a normalizing flow to explore global statistical descriptors of galaxy populations, such as the mean, standard deviation, and histograms of 14 galaxy properties. We find that subhalo gas mass is the most informative feature for constraining the WDM mass, achieving a determination coefficient of R^2 = 0.9. We employ symbolic regression to extract simple, interpretable relations with the WDM particle mass. Finally, we adopt a more localized approach by selecting individual dark matter halos and using a Graph Neural Network (GNN) with a normalizing flow to infer the WDM mass, incorporating subhalo properties as node features and global simulation statistics as graph-level features. The GNN approach yields only a residual improvement over MLP models based solely on global features, indicating that most of the predictive power resides in the global descriptors, with only marginal gains from halo-level information.

Figures (13)

• Figure 1: Histogram of 1024 unique WDM particle masses used in the varied cosmology simulation suite, ranging between 1.8 keV and 16 keV, sampled uniformly from an inverse distribution.
• Figure 2: Two graphs examples built from galaxy catalogs (two different DREAMS-IllustrisTNG simulations). The purple nodes represent the galaxies of the halo and they are connected if their distance is smaller than the linking radius $r_{\rm{link}}$.
• Figure 3: Architectures used for inferring the WDM particle mass from the DREAMS simulations. Left panel: the input consists of global statistical descriptors of galaxy properties across the simulation box, such as the mean, standard deviation, or histogram bins. These features pass through multiple perceptron layers (two are shown for illustration), producing a latent representation that serves as input to the normalizing flow. Right panel: the graphs are constructed from the subhalos hosted by the halos. These graphs pass through multiple graph layers (two are shown), which produce a latent space representation. In both cases, the resulting latent vector is passed to the normalizing flow, which outputs the posterior distribution of the WDM mass conditioned on either the global descriptors or the halo-specific graph.
• Figure 4: Feature importance analysis for predicting the warm dark matter (WDM) mass using histograms of various galaxy properties from all subhalos within a simulation box. The left panel displays the $R^2$ values for the inference. The right panel illustrates the shape of individual sample inferences from a selected simulation, highlighting the constraining power of $M_g$, $M_{\text{tot}}$, and $V_{\max}$. The dashed black line marks the true value of WDM mass for this chosen simulation.
• Figure 5: $1/m_{WDM}$ predictions (y-axis) using the MLP + NF model, compared to the true $1/m_{WDM}$ value of the WDM particle mass used in each simulation (x-axis). A one-to-one red line is shown to indicate where predictions match the true values. The left panel shows the results considering the 14 galaxy properties; the middle and right panels use only the gas mass $M_{g}$ and total mass $M_{t}$, respectively, as single features. Error bars correspond to the 68% credible interval, computed as the [50th–16th, 84th–50th] percentiles of the posterior distribution.