Eliot Ayache
As a data-driven science, Astrophysics has tremendous potential as both driver and consumer of machine-learning (ML) applications. In particular, astrophysicists glean the greatest physical insight from large-scale, population-based studies, where ML techniques can be used to great advantage for dimensionality reduction, generative modeling, and inference.
Recent work ML has made significant strides in both image-domain and time-series techniques. We present a case study using variational autoencoders (VAEs), applied to X-ray observations of Gamma-ray bursts (GRBs) from the Neil Gehrels Swift Observatory. We explore prospects for obtaining physical inferences from ML models. We find that the observed diversity in Gamma-ray burst X-ray emission arises naturally from the competition between a few (2-3) independent physical processes, which produce distinct peaks at different times in the data.
Our work unifies the prevalent classification of GRBs based on their X-ray data into a single continuum, opening new avenues for physical insight into the processes producing the most powerful explosions in the Universe. VAEs and related techniques have the potential of operating as workhorses during upcoming high-cadence time-domain surveys such as the Vera Rubin Observatory and the Square Kilometer Array.
NASA’s Swift satellite has found hundreds of such X-ray afterglows, and every single one of them looks different. The wide variety of X-ray light-curves (the evolution of brightness over time) has baffled astronomers. To better understand this diversity, astronomers have attempted to classify Gamma-ray bursts into categories based on the shapes of their X-ray light curves.
The X-ray brightness typically fades as a power law, or a series of broken power laws, over time. The traditional X-ray classification relies on the number of breaks in these power laws and the relative steepness of the power law slopes between these breaks. The resultant classification has 6 categories. However, the physical origin (and relevance) of the categories is still an open question. Could they be the results of a human bias or do these light-curves cluster together by shape when stripped to their most significant features?
Extracting the most significant features is called “dimensionality reduction”. Here, we use artificial neural networks to “learn” these features. We give a light curve as input to a first network, the encoder, which reduces the light curve to only two numbers. These numbers are then fed to a second network (the decoder) that attempts to regenerate the entire light-curve. We compare the output light curve to the input and adjust the parameters of the encoder and decoder to find the best compressed representations of the light-curves in two dimensions. We expect that if GRB X-ray light curves are truly drawn from categories characterised by specific light curve shapes, then these categories should show up as clusters in the compressed two-dimensional representation.
After having trained the networks, we make the following observations: First, our machine-learning model reproduces all the input light curves well, confirming our ability to reduce the dimensionality of the problem using this technique. Second, clusters from each category overlap significantly in this two-dimensional space. The classes are not well-separated in terms of the principal features of the light curves, as described by the compressed representation. We show that this is not due to the choice of model, as we are able to find clusters in synthetic data directly built from the traditional classes. As a final test of the traditional classification, we encode the real data using a machine learning model pre-trained using the synthetic data. Whereas we expect the real data to follow the clustered distribution of the synthetic data, we again find no clear clusters. This suggests that the real data does not follow the traditional classification schema as well as previously thought.
Our machine learning algorithm is a generative model. This allows us to obtain physical insight from the distribution of GRBs mapped to the compressed space. We compute new decoder outputs (new light-curves) from a grid of points spanning the compressed space, and find that these model light curves exhibit a continuum of morphologies. The central single power-law light curve smoothly transforms into the other shapes, without sharp transitions. Finally, the location of a light curve in the compressed space appears to measure the relative importance of two morphological features: an early spike and a late bump. This finding suggests potential new underlying physical processes, such as the presence of competing emission components, and provides new clues to the origin of X-ray emission in Gamma-ray burst afterglows.