Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning for Gamma-Ray Bursts: A data driven event framework for X/Gamma-Ray analysis in space telescopes

Riccardo Crupi

TL;DR

This work presents DeepGRB, a data-driven framework for high-energy transient detection in space-based GRB monitors, anchored by a neural-network background estimator and the FOCuS-Poisson anomaly detector. It demonstrates how to process archival GBM data to confirm known GRBs and reveal new candidates, with localization, duration, and classification estimates augmented by XAI analyses. The study surveys GRB phenomenology, instrument ecosystems, and AI methods, then integrates background modeling, online triggering, and automatic classification into a cohesive pipeline tailored for next-generation missions like HERMES Pathfinder. The results highlight both the promise of data-driven approaches to uncover faint, long transients and the importance of explainability to diagnose biases and guide future improvements in GRB detection and analysis. Overall, DeepGRB offers a scalable, interpretable framework for on-board and offline analyses that can extend to future missions and multi-instrument data streams.

Abstract

This thesis comprises the first three chapters dedicated to providing an overview of Gamma Ray-Bursts (GRBs), their properties, the instrumentation used to detect them, and Artificial Intelligence (AI) applications in the context of GRBs, including a literature review and future prospects. Considering both the current and the next generation of high X-ray monitors, such as Fermi-GBM and HERMES Pathfinder (an in-orbit demonstration of six 3U nano-satellites), the research question revolves around the detection of long and faint high-energy transients, potentially GRBs, that might have been missed by previous detection algorithms. To address this, two chapters introduce a new data-driven framework, DeepGRB. In Chapter 4, a Neural Network (NN) is described for background count rate estimation for X/gamma-ray detectors, providing a performance evaluation in different periods, including both solar maxima, solar minima periods, and one containing an ultra-long GRB. The application of eXplainable Artificial Intelligence (XAI) is performed for global and local feature importance analysis to better understand the behavior of the NN. Chapter 5 employs FOCuS-Poisson for anomaly detection in count rate observations and estimation from the NN. DeepGRB demonstrates its capability to process Fermi-GBM data, confirming cataloged events and identifying new ones, providing further analysis with estimates for localization, duration, and classification. The chapter concludes with an automated classification method using Machine Learning techniques that incorporates XAI for eventual bias identification.

Deep Learning for Gamma-Ray Bursts: A data driven event framework for X/Gamma-Ray analysis in space telescopes

TL;DR

This work presents DeepGRB, a data-driven framework for high-energy transient detection in space-based GRB monitors, anchored by a neural-network background estimator and the FOCuS-Poisson anomaly detector. It demonstrates how to process archival GBM data to confirm known GRBs and reveal new candidates, with localization, duration, and classification estimates augmented by XAI analyses. The study surveys GRB phenomenology, instrument ecosystems, and AI methods, then integrates background modeling, online triggering, and automatic classification into a cohesive pipeline tailored for next-generation missions like HERMES Pathfinder. The results highlight both the promise of data-driven approaches to uncover faint, long transients and the importance of explainability to diagnose biases and guide future improvements in GRB detection and analysis. Overall, DeepGRB offers a scalable, interpretable framework for on-board and offline analyses that can extend to future missions and multi-instrument data streams.

Abstract

This thesis comprises the first three chapters dedicated to providing an overview of Gamma Ray-Bursts (GRBs), their properties, the instrumentation used to detect them, and Artificial Intelligence (AI) applications in the context of GRBs, including a literature review and future prospects. Considering both the current and the next generation of high X-ray monitors, such as Fermi-GBM and HERMES Pathfinder (an in-orbit demonstration of six 3U nano-satellites), the research question revolves around the detection of long and faint high-energy transients, potentially GRBs, that might have been missed by previous detection algorithms. To address this, two chapters introduce a new data-driven framework, DeepGRB. In Chapter 4, a Neural Network (NN) is described for background count rate estimation for X/gamma-ray detectors, providing a performance evaluation in different periods, including both solar maxima, solar minima periods, and one containing an ultra-long GRB. The application of eXplainable Artificial Intelligence (XAI) is performed for global and local feature importance analysis to better understand the behavior of the NN. Chapter 5 employs FOCuS-Poisson for anomaly detection in count rate observations and estimation from the NN. DeepGRB demonstrates its capability to process Fermi-GBM data, confirming cataloged events and identifying new ones, providing further analysis with estimates for localization, duration, and classification. The chapter concludes with an automated classification method using Machine Learning techniques that incorporates XAI for eventual bias identification.
Paper Structure (96 sections, 1 theorem, 59 equations, 92 figures, 17 tables)

This paper contains 96 sections, 1 theorem, 59 equations, 92 figures, 17 tables.

Table of Contents

  1. Abstract
  2. Acknowledgement
  3. Introduction
  4. State of the art - GRB
  5. GRB overview
  6. Prompt phase
  7. Afterglow
  8. Progenitors
  9. GRB model
  10. Ultra long-GRB
  11. Open questions
  12. Telescopes for GRB
  13. BATSE
  14. BeppoSAX
  15. Swift
  16. ...and 81 more sections

Key Result

Theorem 1

Define a set of samples from a distribution denoted as $Y = {y_1, y_2, \ldots, y_n}$, a constant estimate $\hat{z}$. To minimise the MSE $\frac{1}{n}\sum_{i=1}^{n}(y_i - \hat{z})^2$, $\hat{z}$ must be the mean $\sum_{i=1}^{n}(y_i)$. To minimise the MAE $\frac{1}{n}\sum_{i=1}^{n} \mid y_i - \hat{z} \

Figures (92)

  • Figure 1: My three rabbits, from left to right: Nuvola, Olaf and Arturo.
  • Figure 2: The discovery of GW170817 and GRB 170817A using data from Fermi-GBM and INTEGRAL, along with a time-frequency map from LIGO detectors. abbott2017gravitational © AAS. Reproduced with permission.
  • Figure 3: T90 distribution for GRBs detected with a) BATSE meegan1996third b) Fermi-GBM von2020fourth © AAS. T90 is the time at which the cumulative GRB counts increase from 5% to 95% of the total detected counts. Reproduced with permission.
  • Figure 4: Examples of GRBs lightcurves pe2015physics. The lightcurves of GRB are extremely diverse, with few recognizable patterns. This sample includes short and long events, single peaks or multiple peaks, noisy or very smooth, symmetric asymmetric profiles. Data from the public BATSE archive (http://gammaray.msfc.nasa.gov/batse/grb/catalog/), credict to Daniel Perley.
  • Figure 5: Three elemental spectral components that shape GRB prompt emission spectra: (I) a non-thermal Band-function component, (II) a black-body (quasi-thermal) component, and (III) an additional power-law component. The Fermi satellite utilizes two instruments, GBM (Gamma Burst Monitor) and LAT (Large Area Telescope), each with different energy sensitivities. zhang2011comprehensive © AAS. Reproduced with permission.
  • ...and 87 more figures

Theorems & Definitions (1)

  • Theorem 1