Towards characterizing dark matter subhalo perturbations in stellar streams with graph neural networks

Ping Ma, Keir K. Rogers, Ting S. Li, Renée Hložek, Jeremy J. Webb, R. Q. Huang, Jean‐Luc MeunierarXiv 2025 · DOI · arXiv

This paper, recently published by Ma et al., represents a significant leap in our ability to hunt for the "invisible" building blocks of our universe: dark matter subhalos. By moving away from traditional 1D signal processing and embracing Graph Neural Networks (GNNs), the researchers have unlocked a much more sensitive way to detect the gravitational fingerprints left by dark matter on the Milky Way’s stellar streams.

TL;DR

Researchers used Graph Convolutional Neural Networks (GCNNs) and Simulation-Based Inference to detect dark matter subhalos by analyzing the 6D "phase space" (positions and velocities) of stars in stellar streams. This approach improves mass constraints on these invisible clumps by up to a factor of 11 compared to current state-of-the-art methods, proving that how we look at the data is just as important as how much data we have.

What Problem Does This Paper Address?

The Cold Dark Matter (CDM) model—our best guess for how the universe is structured—predicts that the Milky Way should be swarming with thousands of tiny dark matter "subhalos." These clumps are too small to host stars, making them effectively invisible. However, they aren't totally ghost-like; they have gravity.

When one of these subhalos passes through a stellar stream (a long, thin ribbon of stars torn from a cluster), it acts like a gravitational bowling ball, creating gaps, "spurs," and velocity kicks in the stream. For years, astronomers have tried to find these subhalos by looking at the 1D density power spectrum—essentially looking for "wiggles" in the density of stars along the stream. The problem? This method ignores a massive amount of information, specifically the 3D velocities of the stars and their complex spatial arrangements, making it difficult to distinguish between a small, slow subhalo and a large, fast one.

Methods

The authors developed a pipeline that combines two cutting-edge machine learning techniques:

  1. Graph Convolutional Neural Networks (GCNNs): Instead of treating a stellar stream like a 1D line or a 2D image, the GCNN treats it as a graph. Each star is a "node," and its nearest neighbors are connected by "edges." This allows the network to learn the local geometric and physical relationships between stars in full 6D phase space (3D position + 3D velocity).
  2. Simulation-Based Inference (SBI): Because we don't have a simple formula for the "likelihood" of a subhalo interaction, the authors used thousands of simulations to "teach" the model. They used Masked Autoregressive Flows (a type of generative model) to turn the GCNN’s compressed data into a posterior distribution—a probabilistic map of the subhalo’s most likely mass and velocity.

Datasets

The study relied on synthetic data generated using the streamspraydf algorithm (part of the galpy package).

  • Training Set: Between 40,000 and 80,000 simulations of a "GD-1-like" stream (modeled after one of the most famous streams in our galaxy).
  • Parameters: They varied subhalo masses from $10^5$ to $10^{10}$ solar masses ($M_\odot$) and relative velocities from 22 to 1100 km/s.
  • Benchmarks: The GCNN was tested against a "standard" baseline: a dense neural network trained on the traditional 1D density power spectrum.

Key Insights and Ideas

  • Breaking the Mass-Velocity Degeneracy: In 1D density maps, a fast, heavy subhalo looks nearly identical to a slow, light one. By including velocity data, the GCNN can "see" the specific impulse given to stars, allowing it to tell these two scenarios apart.
  • Quality Over Quantity: One of the most striking findings was that a survey of just 300 stars with full 6D data (position + velocity) provides constraints as strong as a survey of 3,000 stars with only 3D data (position only). This suggests that future missions should prioritize high-precision spectroscopy and astrometry over simply finding more stars.
  • Coordinates Matter: The model performed significantly better when the data was presented in heliocentric coordinates (as seen from Earth) rather than galactocentric coordinates. This is likely because the "spurs" and "gaps" created by subhalos are more visually distinct from our specific vantage point.

Key Results

  • Massive Sensitivity Boost: The GCNN improved constraints on subhalo mass by factors of 11x (for $10^8 M_\odot$), 7x (for $10^7 M_\odot$), and 3x (for $10^6 M_\odot$) compared to the power spectrum method.
  • Better Calibration: The GCNN didn't just give smaller error bars; it gave more honest ones. The "Simulation-Based Calibration" tests showed that the GCNN's predicted probabilities matched the true distribution of errors much better than the traditional method.
  • Information Extraction: The GCNN was able to extract useful information even from very weak interactions where the 1D power spectrum saw nothing but noise.

Limitations and Future Work

  • The "One-Hit" Assumption: This study assumed a single subhalo interaction. In reality, a stellar stream might be bombarded by dozens of subhalos over its lifetime.
  • Baryonic "Noise": Real streams are also perturbed by giant molecular clouds and the Milky Way’s central bar. Future models will need to distinguish these "normal" gravitational kicks from dark matter signals.
  • Observational Realism: The simulations used here were "clean." The next step is to incorporate "selection functions"—the reality that telescopes miss certain stars or have specific measurement errors.

The authors conclude that GCNNs are the future of stream analysis, providing a path to finally "see" the smallest, darkest components of our galaxy.

Summary updated

Abstract

The phase space of stellar streams is proposed to detect dark substructure in the Milky Way through the perturbations created by passing subhalos - and thus is a powerful test of the cold dark matter paradigm and its alternatives. Using graph convolutional neural network (GCNN) data compression and simulation-based inference (SBI) on a simulated GD-1-like stream, we improve the constraint on the mass of a [$10^8$, $10^7$, $10^6$] $M_\odot$ perturbing subhalo by factors of [11, 7, 3] with respect to the current state-of-the-art density power spectrum analysis. We find that the GCNN produces posteriors that are more accurate (better calibrated) than the power spectrum. We simulate the positions and velocities of stars in a GD-1-like stream and perturb the stream with subhalos of varying mass and velocity. Leveraging the feature encoding of the GCNN to compress the input phase space data, we then use SBI to estimate the joint posterior of the subhalo mass and velocity. We investigate how our results scale with the size of the GCNN, the coordinate system of the input and the effect of incomplete observations. Our results suggest that a survey with $10 \times$ fewer stars (300 stars) with complete 6-D phase space data performs about as well as a deeper survey (3000 stars) with only 3-D data (photometry, spectroscopy). The stronger constraining power and more accurate posterior estimation motivate further development of GCNNs in combining future photometric, spectroscopic and astrometric stream observations.
BibTeX
@misc{ma2025towards,
  author = {Ping Ma and Keir K. Rogers and Ting S. Li and Renée Hložek and Jeremy J. Webb and R. Q. Huang and Jean‐Luc Meunier},
  title = {{Towards characterizing dark matter subhalo perturbations in stellar streams with graph neural networks}},
  year = {2025},
  eprint = {2502.03522},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.03522},
  howpublished = {arXiv},
}
Figure