Dark Matter and Stellar Stream Detection

Stellar stream and dark matter detection are active areas of research in astrophysics. Historically, this detection has been a manual process and has led to the discovery of stellar streams such as GD-1 and Palomar 5. If astrophysicists initially have strong confidence that certain stars are part of a stellar stream, probabilistic methods may determine the likelihood that an adjacent star also belongs to the same stellar stream. Using machine learning methods, we seek to develop an automated process that finds additional stars that are part of the stellar stream. After initially trying standardized classification methods, we establish an ensemble of a subset of kNN methods (ESkNN) as an effective means of classifying stars from simulated streams and real streams. Since these methods generally fail to capture non-linearities in the data, we also use a feedforward neural network approach that we expected to help capture non-linear properties. Our evaluation metric for these two metrics is the F1 score. The best model produced by ESkNN method had an F1 score of 0.75, when evaluated on GD-1 stars. The deep learning method (DeepSets) achieved an F1 score of 0.66. While both approaches can be tuned to prefer precision or recall, the ESkNN tends towards highly precise results (99%) whereas the DeepSets method tends towards results with high recall. We also attempt a metric learning approach that could help uncover a distance learning metric that is an improvement upon Euclidean distance.

A paper based on an extension of this work has been submitted to NeurIPS 2020.