Machine learning for electron PID in ATLAS

ATLAS identifies electrons using a likelihood discriminator, as this is a general and transparent approach. However, machine learning algorithms have the ability to use the correlations between the variables and their differences between signal and background to separate these further. Unfortunately, the ATLAS simulation is not accurate enough to use for detailed training, and most of the gain from ML algorithms trained on simulation is lost in these differences, when applied to data. Tag&Probe selection of elections from Z->ee events typically provide samples of 80-95% purity, but using the independency of some variables on the probe electron, this has been increased to above 99%, allowing ML training on data. The resulting ML separation is a clear improvement, which varies significantly (and surprisingly) with energy and direction.