BTSbot · AIxSci

What it is

A telescope that scans the whole visible sky repeatedly will see something change in brightness in millions of places. Each change is sent out as an alert: a small package holding cut-out images of the spot before, during and after the change, plus measurements of how bright it is. Most alerts are not real events at all, but artefacts of the camera or the processing. Of the real ones, most are not exploding stars but flickering galaxy centres, restless variable stars or close pairs of stars trading matter. BTSbot is a neural network trained to read these packages and say which ones look like a bright supernova, an exploding star near enough to study in detail.

A neural network is a mathematical function with many adjustable numbers inside it. Nobody writes the rules by hand. Instead it is shown a large collection of examples already labelled by people, and the numbers are nudged over and over until its answers match those labels. For this kind of task the examples are the alert images themselves, sorted into classes such as confirmed bright supernovae, flickering galaxy centres and variable stars, and candidates that human scanners looked at and chose not to keep. Once trained, the network scores each new alert in a fraction of a second, and it sits inside a longer filter alongside ordinary hand-written rules.

Why scientists use it

The bottleneck in this kind of survey is not finding candidates but choosing among them. A supernova is most informative in its first hours, when the blast wave is still ploughing through whatever the star shed before it died. Deciding which of the night's alerts deserve a spectrum, a measurement that splits the light into its colours to reveal what the object is made of, used to mean people scrolling through image stamps every morning. A trained model does that sorting continuously, with the same criteria every night, which means a request for follow-up observations can be issued while the object is still rising.

In the record

astronomyNeural network picks bright exploding stars from sky survey alerts and books telescope timeTrained on labelled ZTF alert packets covering confirmed bright transients, galaxy nuclei and variable stars, dim transient-like sources, and candidates human scanners declined to save.astronomySoftware spotted a nearby supernova and booked a telescope within minutesWorked inside a mostly rule-based real-time filter, after a separate network scored alerts as real or spurious with its pass threshold raised to suppress false alerts over bright galaxy centres.astronomyAstronomers test pre-trained image models for sifting telescope alertsServed as the baseline model when its training set was rebuilt and compared against networks pre-trained on survey colour images, including runs using only small fractions of the alerts.

Limits

Such a model learns the judgements it was shown, including the quiet preferences of the people who built its training set, so it can miss the odd events that no label describes. It is tuned for bright, nearby objects and says little about faint ones. It cannot explain its scores, and it depends on earlier filtering steps: when the real-or-spurious model before it is loosened or tightened, what reaches the classifier changes. It sorts candidates; it does not measure anything about the star.

Written by AIxSci: the background from general knowledge, the examples from checked records.