Title:Reconstruction of Gamma Ray Burst Lightcurves: A Machine/Deep Learning Approach

Abstract:Gamma-ray bursts (GRBs), which are highly energetic bursts, have been detected at high redshifts (z = 9.4) and are pivotal for cosmological research and the study of Population III stars. To address these studies, it is important to identify correlations between the key GRB variables while minimizing uncertainties. This requires comprehensive coverage of GRB Light curves (LCs), which are challenging to obtain given the observational gaps and noise. This study explores the reconstruction of GRB LCs using three machine-learning models: Bi-LSTM, CGAN, and a Sarimax-based Kalman filter. To evaluate these approaches, we also include well-established functional form parametric methods such as the Willingale model (W07) and the non-parametric Gaussian Processes (GPs), expanding upon arXiv:2305.12126 by applying our techniques to a larger sample of 545 GRBs, covering four different GRB classifications. Our findings reveal that the LSTM model achieves the highest reductions in the uncertainties of the plateau parameters, with decreases of 29% in log Ta, 31% in log Fa and 40% in {\alpha}. In comparison, GPs show a slightly smaller reduction of 22% in both log Ta and log Fa, and 32% in {\alpha} while CGANs exhibit a decrease of 23% in log Ta, 24% in log Fa, and 31% in {\alpha}. The Sarimax-based Kalman filter performs well for well-sampled and flare-free GRBs, but it has problems with reconstruction in more complex cases. It shows a decrease of 14% in log Ta, 5% in log Fa and 6% in {\alpha}. [..] For the subset of 299 Good GRBs, namely with well-defined plateau emission with no flares and bumps, the LSTM model achieved a reduction of approximately 40%, with a percentage increase of 8% compared to the results observed in arXiv:2305.12126. [..] These advancements are essential for using GRBs as standard candles in cosmology and investigating theoretical models.[..]