Uncovering dark matter density profiles in dwarf galaxies with graph neural networks
Interdisciplinary Research Achievement �This project introduces a novel approach for inferring the dark matter distribution in dwarf galaxies using observed kinematics of stars bound to these systems using graph-based machine learning. This aims to address several limitations of established methods and can place stronger constraints on dark matter profiles in these systems, with the potential to resolve some of the ongoing puzzles associated with the small-scale structure of dark matter. A paper has been published [1] detailing and validating the neural simulation-based inference method and code has been made public [2]. Additionally, progress has been made on validating the method on realistic (hydrodynamic) simulations from the FIRE suite in order to understand its out-of-distribution performance compared to traditional methods used so far. This will form a foundation for applying it to real data.
Impact on Artificial Intelligence�The method combines, for the first time, graph-based machine learning (specifically, graph-convolutional neural networks) with simulation-based inference. Although the method is applied to the physical system of dwarf satellite galaxies, it is a general approach that can be used to perform simulation-based inference using point cloud data of any kind.
Impact on Fundamental Interactions
The specific format in which observations are available---the 2-D angular coordinates of the stars, and the 1-D velocity along the line of sight measured using spectroscopy---means that this structure needs to be carefully accounted for both in constructing the graph out of stellar kinematic data, as well as in designing the graph neural network. Imposing this structure as an inductive bias was found to be necessary in order to enable a simulation-efficient analysis. Ultimately, the goal is to apply the method to real observations of dwarf galaxies in order to infer their dark matter profile and interpret the result in terms of possible dark matter microphysics scenarios, such as self-interacting dark matter, which is of considerable interest to the particle physics community.
Tri Nguyen (MIT), Siddharth Mishra-Sharma (IAIFI Fellow), Lina Necib (MIT)
The NSF Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) is �supported by National Science Foundation under Cooperative Agreement PHY-2019786
Schematic illustration of the method
Outlook & References
The next steps will be to robustly understand how well the method performs on realistic, hydrodynamical simulations and if necessary, improve the method for better out-of-distribution generalization. Then, we will apply the method to real dwarf galaxy data to uncover the latent dark matter distribution and understand the consequences of the results for particle physics models like self-interacting dark matter.
[1] https://journals.aps.org/prd/abstract/10.1103/PhysRevD.107.043015; [2] https://github.com/trivnguyen/JeansGNN