The aim

The aim is to alert on the most interesting classes of variable objects so that follow-ups using ground-based telescopes can begin. This includes supernovae, microlensing events, near-Earth asteroids, novae, stars undergoing rare and interesting phases of evolution (such as helium flash), and so on. Every class of object will require an individual trigger.

The number of alerts

object	4 years	per day	source
Supernovae	6800	4	Belokurov and Evans 2003

A prototype built with Self-Organizing Maps

Read the full invited review for "The Three Dimensional Universe with GAIA".

SOMs are a powerful way to take a quick-look at the data. They provide a broad brush clustering of the main types of pattern very quickly. They are ideal for Petabyte datasets (like GAIA). This is because they are fast, unsupervised and make no prior assumptions about the data.

SOMs are a powerful way of implementing novelty detection. The number of nodes is roughly the number of distinct classes and is always much smaller than the number of different patterns. The nodes represent the most abundant patterns. If a pattern is rare, then necessarily there will be no node allocated near to it. So, such patterns are identifiable through their large ``quantisation error''. Cuts on the quantisation error and the signal-to-noise ratio can substantially reduce the amount of data. This may already be enough to identify the high-quality discrepant patterns on which we wish to alert.

1. Every 6 hours, a SOM is built from the GAIA datstream. The discrepant patterns are extracted with cuts on signal-to-noise and quantization error. Also extracted are the common patterns corresponding to known types of variable stars.
2. The discrepant patterns are cross-checked against a catalogue of known stellar variables. Some of the stellar variables can be pre-loaded from existing surveys of variable stars (such as those available from the microlensing surveys). This catalogue however will be incomplete at the beginning of the mission and so will need to be up-dated every 6 hours with new variables identified by the SOM.
3. If the discrepant patterns are not in the catalogue of variable stars, then they are candidates for alerts and need to be looked at very closely. It may be that there is already enough confidence that the object needs ground-based follow-up to issue an alert. It may be that further tests are needed for specific classes of object.

More on lightcurve classification with SOMs