Machine Learning Identification of Gravitationally Microlensed Gamma-Ray Bursts

TL;DR

The study addresses identifying gravitationally microlensed gamma-ray bursts (GRBs) in large datasets by training classifiers on simulated lensed/nonlensed GRB light curves. It leverages Cesium-based light-curve features and compares six classifiers, with Random Forest delivering the best performance (≈86% accuracy, AUC ≈0.94) on synthetic data. A simulation-to-reality gap is observed when applying models to six real Fermi/GBM candidates, driven by distributional differences in features between simulated and real GRBs. The work demonstrates ML-based microlensing identification is feasible and highlights the need for improved simulations and robust features to enable reliable real-time candidate detection.

Abstract

Gravitational microlensing of gamma-ray bursts (GRBs) provides a unique opportunity to probe compact dark matter and small-scale structures in the Universe. However, identifying such microlensed GRBs within large data sets is a significant challenge. In this study, we develop a machine learning (ML) approach to distinguish lensed GRBs from their nonlensed counterparts, using simulated light curves. A comprehensive data set is generated, comprising labeled light curves for both categories. Features are extracted using the Cesium package, capturing critical temporal properties of the light curves. Multiple ML models are trained on the extracted features, with Random Forest achieving the best performance, delivering an accuracy of 86% and an F1 score of 0.86 (0.87) for the nonlensed (lensed) class. This approach successfully demonstrates the potential of ML for identifying gravitational lensing in GRBs, paving the way for future observational applications.

Figures (9)

• Figure 1: Left: simulated lensed light curve (see Eq. \ref{['eq:superimpose_lensed']}); Right: simulated non-lensed light curve (see Eq. \ref{['eq:superimpose_not_lensed']}).
• Figure 2: Pairwise feature comparison of the extracted features. The lower triangle displays scatter plots with linear regression fits (red lines) and annotated Pearson correlation coefficients ($r$). Diagonal panels show kernel density estimates (KDEs) representing the distribution of each feature.
• Figure 4: Normalized confusion matrices for six machine learning models used to classify lensed and non-lensed GRBs.
• Figure 5: ROC curves for all models. The area under the curve (AUC) for each model is indicated in the legend. The dashed blue line indicates a model with an AUC of 0.5, equivalent to random guessing with no discriminatory ability.
• Figure 6: Light curves of selected sources from the test set. Each subplot shows the time series data for a specific source, with the true label (Lensed or Non-Lensed) and the predicted label from our machine learning model. The color coding distinguishes between Non-Lensed (red) and Lensed (blue) sources.