Explainability of Unsupervised Machine Learning Methods for High-Energy Astrophysics

Gamma-ray bursts (GRBs) are the most luminous bursts of light in our universe. They are grouped as long or short based on the events that caused them. Short GRBs tend to be a product of binary neutron star mergers and extragalactic magnetar giant flares, and long GRBs can come from collapsars, a rare type of core-collapse supernovae. The goal of this project is to increase our understanding of unsupervised dimensionality reduction algorithms, using GRB data. Our algorithm is trained through waterfall plots. The waterfall plots are a data product that uses the GRB prompt emission data to characterize events captured by Fermi-GBM. They contain the complete set of spectral, temporal, and correlated information available in our observational data. We utilized the embedding produced with convolutional autoencoders as presented in Negro et al 2025, which is a trusted 30-dimensional representation of the input dataset. This is processed through UMAP, which is a dimensionality reduction technique, that reduces the 30-dimensional space to two and three dimensions. The focus of this research is to optimize this algorithm and produce a reliability score for the GRB representation in lower dimensions. This study is crucial to reliably interpret the GRB embedding and rank follow up priorities for other facilities to observe the most interesting GRBs.