Speaker
Description
Astronomical transients, including Fast Radio Bursts and Gamma-ray Bursts, are among the most energetic known phenomena in the universe, originating mostly from extragalactic distances. To gain deeper insight into their origin and nature, it is essential to analyse their substructure. A rapidly increasing number of observations allows for data-driven analysis of these transients. In this talk I will present a machine learning (ML) model able to analyse the sub-burst structure of astronomical transients. Light curves of transients often consist of several components and can thus be modeled as a superposition of multiple short sub-bursts. Each of these sub-bursts can be characterised by several parameters, including their arrival time and skewness, which carry important information about the trigger and emission processes. One of the main challenges is that the number of sub-bursts within each transient time series is unknown, making traditional sampling methods non-trivial to implement and computationally inefficient. To address this, our ML model first uses an encoder followed by a multi-layer perceptron to infer a distribution over the most likely number of sub‑bursts. Using this information, a transformer-based model then infers the parameter posteriors of each sub-burst using flow matching. To gain insight into the robustness and performance of the model, we apply it to both synthetic and real Fast Radio Bursts (FRBs), which are very bright and short millisecond duration radio transients, observed by the Canadian Hydrogen Intensity Mapping Experiment telescope (CHIME).