AGILE is a space mission launched in 2007 to study X-ray and gamma-ray astronomy. The AGILE Team is developing new detection algorithms for Gamma-Ray Bursts (GRBs) both with classical and machine learning technologies. To train or test these algorithms, it is necessary to have a large GRB dataset, but usually, there are not enough real data available. This problem is also common for the new generation of high-energy astrophysics projects (such as COSI and CTA). It therefore becomes essential to have a system to simulate GRB data.
This work aims to develop a Deep Learning-based model for generating synthetic GRBs that closely replicate the properties and distribution of real GRBs. The dataset obtained using the trained model can then be used to develop detection algorithms both with classic techniques and with machine learning. To develop this generative model we have to take into account several complexities. The main ones are the huge different temporal behaviours of such GRBs and the lack of data making the training harder.
We propose a new method for generating GRBs. This model combines the Generative Adversarial Network (GAN) and Variational Autoencoder (VAE) to produce high-quality and structured generative results while preserving latent space structure. We also developed a loss function tailored to our dataset for the reconstruction task. To create the training dataset we extracted the light curves (LC) of GRBs presented in the Fourth Fermi-GBM Catalog using only long GRBs. This catalogue contains more than a decade of observations with a total of 3608 GRBs, captured through 12 NaI and 2 BGO detectors. In this work, we considered only the NaI detectors. We used the LCs detected by multiple sodium detectors and related to the same GRB as independent, to augment the number of samples. We filtered the LCs by removing outliers and those with missing values, ending with a total of 5964 LCs. Evaluating the distribution of the GRB duration (through the t90 parameter), we set the length of the time series at 220.
We assessed the model's performance by quantifying the dissimilarity between the histograms of count rates in synthetic and real GRBs. Additionally, we conducted quantitative analysis, employing statistics of the LCs distribution in both datasets. The results show that the synthetic LCs are generated with very similar properties to the real ones. We are working on a conditional version of our model using physical parameters of the GRBs such as the t90 or the fluence to have a more precise generation. This method can be used to generate synthetic LCs for other high-energy astronomy projects such as AGILE, CTA and COSI.