Uncovering Anomalies in Gamma-Ray Bursts X-Ray Afterglows

A great diversity of the Gamma-Ray Bursts (GRBs) prompt emission together with a relatively small number of detected events (several thousand) significantly limit the use of statistical tools in distinguishing between the “ordinary” and “anomalous” objects. In contrast, the X-Ray afterglow emission exhibits a high universality, which makes the statistical approach possible.

However, the present-day analysis of GRB X-Ray afterglow lightcurves is highly model-dependent and probably oversimplified, which obstructs a robust anomaly detection. Here we address this issue and develop a number of model-independent anomaly detection techniques utilizing Machine Learning and Deep Learning models.

In a few words, we are looking for such model that:

• encodes an X-Ray lightcurve in terms of 2-3 features, which correlate with the lightcurve morphology (e.g. with the number of breaks)
• estimates the anomaly measure for each lightcurve and, using this measure, detects at least conventionally anomalous GRBs such as GRB 221009A (its X-Ray afterglow lightcurve together with the corresponding pseudo-color timelapse animation, both measured by Swift-XRT, are shown in the figure above)

We demonstrate that AutoEncoder is a promising model satisfying these criteria.

Inference

The main product presented by this project is a pre-trained AutoEncoder-based GRB X-Ray Afterglow Anomalous Light Curves Detector employing swifttools API. The detector works in a user-friendly “black-box” manner, without requiring the user to understand what is going on in detail, but still being flexible enough for those who wants to manage the parameters.

inference/README.md presents a complete User's Guide for this detector.

Dataset

This study utilizes the Swift-XRT GRB lightcurve [repository] data. The dataset is briefly described in dataset/README.md

We focus only on the basic lightcurve entries: the timestamps and source count rate, together with the errors of the latter. We develop and use four preprocessing methods, see dataset/README.md

Models

We analyse lightcurves using Density-based spatial clustering of applications with noise (DBSCAN), IsolaitonForest, Kernel principal component analysis (KernelPCA), and AutoEncoder. Each model is accompanied by its own pipeline of pre-/postprocessing steps. See models/README.md for details.

Tested Environment

Operating System	Python Version	IDE/Editor	Hardware
macOS Sequoia 15.0.1	3.11.5	Jupyter Notebook 6.4.13	MacBook Air (M1, 8GB RAM)

Before trying to run anything on your device, please pay attention to the fact that:

• directories in this repository contain their individual requirements.txt files,
• list of packages provided in each requirements.txt file assumes that you have already installed python and Jupyter Notebook, as well as the corresponding dependencies on your device.

Setup with Conda (recommended)

Please consider installing [miniconda3] and working within individual [environments]. Following the instructions given below will ensure reproducibility, assuming that you work with code locally on your device.

Step 1. Create a new python 3.11.5 environment:¹

$ conda create -n GRB_env python==3.11.5 -y

Step 2. Next, activate the environment:

$ conda activate GRB_env

Step 3. From the current directory (where you read this guide), switch to the folder <directory>:

• inference — to work with a pre-trained anomaly detector based on AutoEncoder model,
• dataset — to create/preprocess dataset,
• models — to work with Machine Learning models,
• models/AutoEncoder — to work with AutoEncoder model,

It is highly recommended to have an individual environment for each of these options. Then install the required packages and Jupyter Notebook 6.4.13 (the latter is optional if <directory> is inference or dataset)

(GRB_env) $ cd ./<directory>
(GRB_env) $ python -m pip install -r requirements.txt [notebook==6.4.13]

Additionally, please ensure that [curl] is installed on your device. This is required for working with inference and models, since the pre-trained weights/prepared datasets, respectively, are by default downloaded from the online version of this repository.

You can now work with:

• inference/utilities python scripts,
• dataset/utilities python scripts,
• DBSCAN.ipynb, IsolationForest.ipynb, and KernelPCA.ipynb notebooks,
• AutoEncoder.ipynb notebook,

respectively. For details, please read <directory>/README.md.

Step 4. When your work is finished, deactivate the environment:

(GRB_env) $ conda deactivate

If you do not need the created environment anymore, you can permanently remove it by:

$ conda remove -n GRB_env --all -y

Contacts

If you have any comments, suggestions, or questions, please send them to martynenko.ns18@physics.msu.ru

Acknowledgements

This work is greatly supported by Non-commercial Foundation for the Advancement of Science and Education INTELLECT. The author is indebted to G.I. Rubtsov, V.A. Nemchenko, and A.V. Ivchenko for helpful discussions.

Footnotes

1. Of course, istead of GRB_env you can assign the environment names at your discretion ↩