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

Overview

The nature of dark matter (DM) remains one of the biggest unanswered questions in physics, all while current methods remain challenged with accurately distinguishing between different DM models. To address this, I developed a simulation-based inference framework that uses the GD-1 stellar stream to test competing DM models and baryonic perturbers. My pipeline generates millions of realistic mock streams and analyzes them with a graph-neural and transformer encoder that combines full phase-space information with density, wavelet, and topological summaries. Thus allowing the model to return calibrated parameter posteriors and probabilities for cold, self-interacting and ultra-light DM, as well as baryonic perturbers. The results robustly favor (Bayes factor >10) DM subhaloes for key features, recover masses to 1σ for M ≥ 5×10⁶ M⊙, and reach sensitivity ≈ 3×10⁶ M⊙. Ultimately delivering quantitative constraints on the Milky Way substructure spectrum, and motivating refined modeling and future multi-probe tests of DM physics.

Awards (2)

• Gold Medal
• Selected for CWSF 2026

Competition history

• CWSF 2026 Aerospace

Related projects

Closest projects by meaning, across every fair and year in the corpus.