Ghosts in the Stream: Probing Dark Matter in the GD-1 Stellar Stream with Simulation-Based Inference

Research Problem: Searching for GD-1’s Missing Mass

Most current astrophysical stream analyses can identify spatial and gravitational perturbations, but do not yet reliably distinguish which object class or dark matter (DM) model caused them, and most lose detection sensitivity as perturber mass decreases.

Motivation:

The uncertainty in perturber class prediction, and the single model assumption most studies take, can lead to degeneracies in interpretation. Thus, limiting our capacity to test DM physics, measure the subhalo mass spectrum, and identify individual perturbers.

Approach:

I focus on a principled Simulation-Based Inference (SBI) pipeline that trains conditional diffusion generative emulators to map physical perturber parameters into realistic GD-1-like mock streams matched to Gaia DR3 and the Price-Whelan and Bonaca member catalogs; amortized emulation gives millions of low-cost forward realizations while a higher-fidelity N-body module validates edge cases.

Project Goal:

Detect and categorize DM subhalos through the dynamical perturbations of the GD-1 stellar stream and achieve explicit model-class discrimination between DM and baryonic perturbers.

Importance:

Develops a classification framework to distinguish between DM and baryonic models.

Extends existing SBI architectures by incorporating novel model-class labels into the neural density estimators.

Moves beyond simple summary statistics by implementing multi-scale and topological descriptors to capture complex stream morphologies.

Integrates Generative Diffusion Models as probabilistic emulators for stellar stream evolution to bypass computationally expensive N-body simulations.

Allows the future integration of gravitational lensing and satellite constraints into the framework.

Scientific Background: Stellar Stream Purturbation Modelling

Model the stream as stars moving along a common orbit in the Milky Way potential.

Add a perturber to the stream (e.g., a dark matter subhalo)

Compute gravitational kick of perturber and advance stream forward in time, creating observable features to study and map.

Quantify the perturbers parameters.

→ Based on perturber’s inferred parameters, we select the best-fitting object class and assign it as the cause of perturbation.

Datasets:

Gaia Data Release 3 (DR3): astrometry and proper motions

GD-1 Member Catalog (Price-Whelan & Bonaca, 2018): membership lists and derived GD-1 stream tracks

Methodology:

Preprocessing

I start by refining GD-1 membership, selecting the highest quality candidates for experimentation. Starting with the initial catalog I refine through astrometric purity, precision cuts, photometric limit, and kenimatic mask before selecting the final sample of GD-1 stars (~0.01% of initial catalog).

Simulation-Based Inference

Setup:

Conditional Emulation

Injection Grid: 10^5-10^8 M⊙ (densest at 10⁶-10^7 M⊙)

Classes: CDM, SIDM, ULDM, Baryons

Neural Density Estimator

10,000 sampler draws/candidate star

Output:

Full Posterior probability of perturber class

Bayes Factors

30% narrower mass posteriors

Calibrated 90% coverage

Novel Conditional Stream Emulation

I then conduct 2,000+ N-body simulations of GD-1-like encounters to forward model millions of mock streams through 4 emulators. These train 400k steps, varying perturber parameters, and validating against high-fidelity N-body cases, with baryonic confuser integration.

Neural Posterior and Classifier Outputs

I use a state-of-the-art 5 GNN layers, 12 attention heads, and 512-dimensional embeddings tested against baryonic mimics and catalog artifacts, to convert stream data into final scientific outputs.

Validation and Calibration

Finally, I test the full pipeline with injection-recovery experiments, ensuring calibration of credible intervals, confidence level, stream morphology, baryonic confusion, and Bayes Factor Stability. Should calibartion fail, I reweight or retrain until outputs are statistically reliable.

Results

Model Probabilities

My novel Simulation-Based Inference pipeline returns a state-of-the-art probability for each dark matter model class.

My results strongly favor Cold Dark Matter (CDM) for the largest perturbations of the GD-1 stellar stream, showing the highest posterior probability.

The highest-mass CDM candidates show the strongest separation from baryonic alternatives, with baryonic probabilities falling to 0.5% or lower for strongest candidates.

Thus, my method showcases a full model-class discrimination, not just simple parameter estimations, as results are consistent with the strong Bayes factors found in the final iteration.

Gap and Power-Spectrum Constraints

My pipeline was able to detect important gaps in the GD-1 stellar stream with depths of 21% - 45% at 2° - 8.5° scales. My method also achieves a state-of-the-art FPR > 5% for scientifically important gaps. Additionally, the smallest recovered gap of 1.9°, shows my method's sensitivity down to compact features.

My strongest detected gap reaches 10.5σ significance.

I managed to constrain the 1-D linear-density Fourier power spectrum in 3 k-bins, with relative power errors of ~0.28 matching the state-of-the-art ±30%

Ultimately, I was able to measure and constrain the GD-1 stream structure at both the gap level and the global power-spectrum level.

Posterior Masses and Uncertainties

My results show that my method's posterior distributions recover the perturber mass directly from the observed GD-1 structure.

Strongest candidates are recovered close to true masse, with median posterior near expected value.

1σ mass recovery for M ≥ 5×10⁶ M⊙, with novel detection sensitivity down to ≈ 3×10⁶ M⊙.

Posterior widths shrinks ~30% from baseline width, showing higher precision.

Lowest-mass candidates remain broader than high-mass ones, but uncertainties support a state-of-the-art model-class comparison.

Thus, showing that stream data can constrain perturber mass at the levels needed to test dark matter substructure predictions.

Bayes Factors

Bayes factors compare how strongly data favors CDM when compared to baryonic perturbers.

Results indicate that all five candidates exceeded BF = 10, with the strongest candidate reaching BF = 140.0

Every candidate passes the robustness requirement of BF/σBF > 5

This shows that my results are not a simple best-fit preference, but a strong statistical discrimination, with evidence becoming progressively stronger for larger perturbers.

Discussion and Conclusion

What this Means for Dark Matter

Results favor DM subhaloes as most likely cause for strongest GD-1 disturbances.

Model attributes highest posterior probability to CDM.

Bayes factors >10 against baryonic explanations.

Stream features successfully distinguish dark matter microphysics.

Results reach sensitivity to ≈ 3×10⁶ M⊙.

Low-mass dark substructures affect stream.

Measured gaps and power-spectrum structure demonstrate that GD-1 can probe subhalo mass spectrums in Milky Way halo.

Thus, GD-1 behaves like gravitational detector for invisible structures; with DM as more plausible explanation than baryonic perturbers for strongest features.

Why Baryonic Confusion still Matters

Not every gap or spur in GD-1 comes from dark matter.

Baryonic perturbers can create stream features similar to dark-matter.

Good fit to the data does not automatically prove dark matter is responsible.

Strongest dark matter claims are safest when strong after testing against baryonic alternatives

Remaining confusion strongest for smaller, shallower, or noisier gaps, where multiple causes look alike

How Calibration Affects Confidence

Calibration tells us whether the model’s probabilities and uncertainties are trustworthy or just look precise. Ultimately, helps confirm that the strong CDM preference is not byproduct of noisy training or poor uncertainty estimates.

Improvements on Current Method

Compares multiple DM and baryonic models directly, adding model-class probabilities and Bayes factors.

Method uses full phase-space, multi-scale density information, and topological features.

Emulators makes forward model faster and more realistic, while keeping uncertainty explicit.

Calibration is stronger than earlier work.

Tests against baryonic confusers, different Milky Way potentials, and held-out N-body cases.

Future Work

Extension to Lensing and Satellites

My project is designed so the same inference framework can later include other DM probes besides GD-1.

Lensing can test small dark objects through optical distortion, creating independent checks on substructure.

Satellite galaxies can help constrain larger DM population by providing another view of how structure forms in the Milky Way halo.

Combining these probes with GD-1 would let analysis compare different lines of evidence vs. only one mode of data. In the long term, this extension could turn my pipeline into a multi-probe framework for testing DM physics.

Special Thanks To:

Pranav Kulkarni (Stanford USA): for mentoring me throughout this process and making time for me in your incredibly busy schedule.

TSF: for all the support in registration, poster design and preparations.

My mom: for supporting me throughout this journey, cheering me on through the ups and down of my project. All of it would not be possible without you!

ÉSTO: À tous mes enseignants et l'administration, merci de m'avoir supporter et encourager à travers cette expérience. Sans vous et notre communauté je n'aurai jamais eu la confiance, la passion ou la détérmination de poursuivre ce projet.

References

Adams, D. K., Parikh, A., Slone, O., Essig, R., Kaplinghat, M., & Price-Whelan, A. M. (2025). Semianalytic Modeling of Dark Matter Subhalo Encounters with Thin Stellar Streams: Statistical Predictions for GD-1-like Streams in Cold Dark Matter. The Astrophysical Journal, 991(1), 66. https://doi.org/10.3847/1538-4357/adf740

Alvey, J. E., Gerdes, M., & Weniger, C. (2023). Albatross: A scalable simulation-based inference pipeline for analysing stellar streams in the Milky Way. Monthly Notices of the Royal Astronomical Society, 525(3), 3662–3681. https://doi.org/10.1093/mnras/stad2458

Bonaca, A., Hogg, D. W., Price-Whelan, A. M., & Conroy, C. (2019). The Spur and the Gap in GD-1: Dynamical Evidence for a Dark Substructure in the Milky Way Halo. The Astrophysical Journal, 880(1), 38. https://doi.org/10.3847/1538-4357/ab2873

Buschmann, M., Kopp, J., Safdi, B. R., & Wu, C.-L. (2018). Stellar Wakes from Dark Matter Subhalos. Physical Review Letters, 120(21). https://doi.org/10.1103/physrevlett.120.211101

de Boer, T. J. L., Erkal, D., & Gieles, M. (2020). A closer look at the spur, blob, wiggle, and gaps in GD-1. Monthly Notices of the Royal Astronomical Society, 494(4), 5315–5332. https://doi.org/10.1093/mnras/staa917

Delos, M. S., & Schmidt, F. (2022). Stellar streams and dark substructure: the diffusion regime. Monthly Notices of the Royal Astronomical Society, 513(3), 3682–3708. https://doi.org/10.1093/mnras/stac1022

Erkal, D., & Belokurov, V. (2015). Properties of dark subhaloes from gaps in tidal streams. Monthly Notices of the Royal Astronomical Society, 454(4), 3542–3558. https://doi.org/10.1093/mnras/stv2122

Erkal, D., Belokurov, V., Bovy, J., & Sanders, J. L. (2016). The number and size of subhalo-induced gaps in stellar streams. Monthly Notices of the Royal Astronomical Society, 463(1), 102–119. https://doi.org/10.1093/mnras/stw1957

Gialluca, M. T., Naidu, R. P., & Bonaca, A. (2021). Velocity Dispersion of the GD-1 Stellar Stream. The Astrophysical Journal Letters, 911(2), L32. https://doi.org/10.3847/2041-8213/abf491

Gilman, D., Birrer, S., Nierenberg, A., & Oh, M. S. H. (2024). Turbocharging constraints on dark matter substructure through a synthesis of strong lensing flux ratios and extended lensed arcs. Monthly Notices of the Royal Astronomical Society, 533(2), 1687–1713. https://doi.org/10.1093/mnras/stae1810

Karouzos, M. (2017). Extreme gravitational lensing. Nature Astronomy, 1(10), 653–653. https://doi.org/10.1038/s41550-017-0281-3

Ma, P. X., Rogers, K. K., Li, T. S., Hložek, R., Webb, J. J., Huang, R., & Meunier, J. (2025). Toward Characterizing Dark Matter Subhalo Perturbations in Stellar Streams with Graph Neural Networks. The Astrophysical Journal, 987(1), 96. https://doi.org/10.3847/1538-4357/add698

Price-Whelan, A. M., & Bonaca, A. (2018). Off the Beaten Path: Gaia Reveals GD-1 Stars outside of the Main Stream. The Astrophysical Journal, 863(2), L20–L20. https://doi.org/10.3847/2041-8213/aad7b5

Schanz, A., List, F., & Hahn, O. (2024). Stochastic Super-resolution of Cosmological Simulations with Denoising Diffusion Models. The Open Journal of Astrophysics, 7. https://doi.org/10.33232/001c.125902

Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., & Poole, B. (2021, February 10). Score-Based Generative Modeling through Stochastic Differential Equations. ArXiv.org. https://doi.org/10.48550/arXiv.2011.13456

Tavangar, K., & Price-Whelan, A. M. (2025). Inferring the Density and Membership of Stellar Streams with Flexible Models: The GD-1 Stream in Gaia Data Release 3. The Astrophysical Journal, 988(1), 45. https://doi.org/10.3847/1538-4357/addd1c

Wagner-Carena, S., Aalbers, J., Birrer, S., Nadler, E. O., Darragh-Ford, E., Marshall, P. J., & Wechsler, R. H. (2023). From Images to Dark Matter: End-to-end Inference of Substructure from Hundreds of Strong Gravitational Lenses. The Astrophysical Journal, 942(2), 75–75. https://doi.org/10.3847/1538-4357/aca525

Zhang, X., Yu, H.-B., Yang, D., & Nadler, E. O. (2025). The GD-1 Stellar Stream Perturber as a Core-collapsed Self-interacting Dark Matter Halo. The Astrophysical Journal Letters, 978(2), L23. https://doi.org/10.3847/2041-8213/ada02b