Late activity of the central engine is often invoked in order to explain the flares observed in the early X-ray afterglow of gamma-ray bursts, either in the form of an active neutron star remnant or (fall-back) accretion onto a black hole. However, these scenarios are not always plausible, in particular when flares are delayed to very late times after the burst. Recently, a new scenario was proposed that suggests X-ray flares can be the result of the passing of a long-lived reverse shock through a stratified ejecta, with the advantage that it does not require late-time engine activity.

In this work, we numerically demonstrate this scenario to be physically plausible, by performing onedimensional simulations of ejecta dynamics and emission using our novel moving-mesh relativistic hydrodynamics code. Improved efficiency and precision over previous work enables the exploration of a broader range of setups. We can introduce a more physically realistic description of the circumburst medium mass density. We can also locally trace the cooling of electrons when computing the broadband emission from these setups.

We show that the synchrotron cooling timescale can dominate the flare decay time if the stratification in the ejecta is constrained to a localised angular region inside the jet, with size corresponding to the relativistic causal connection angle, and that it corresponds to values reported in observations. We demonstrate that this scenario can produce a large range of observed flare times, suggesting a connection between flares and initial ejection dynamics rather than with late-time remnant activity.

We look for hidden patterns in the properties of GRBs. We trained variational autoencoders (fig. 1) on X-ray observations of 1300 GRB afterglows to search for physical patterns in the latent space.

A CNN-based autoencoder provides no evidence for latent-space-based clusters in the X-ray light curves, but correctly clusters a stratified, synthetic training set. Even when trained on synthetic data with well-defined clusters, the CNN model does not yield clusters in the real data, suggesting that the traditional, phenomenological GRB X-ray light curve classes may be somewhat arbitrary. Instead, templates from the latent-space representation of the data reveal that light curve morphology spans a continuum (fig. 2).

There is a large diversity in X-ray afterglow light curves, and no clear evidence for traditional categories when using ML-based clustering methods. The inferred continuum of light curves may be representative of a competition between multiple underlying physical processes.