Title : Optimal graph convolution neural network for efficient particle identification
Supervisor : Frédéric Magniette, Laboratoire Leprince-Ringuet (LLR), CNRS/Ecole
polytechnique, frederic.magniette@llr.in2p3.fr
Funding : project OGCID ANR-21-CE31-0030, 3 years
Location : LLR, Ecole Polytechnique, 91128 Palaiseau Cedex – France

Particle detectors are devices designed to track the evolution of very high speed particle collisions. They produce huge data volume which are energy measures associated to geometric positions. The high energy physics try to find very rare events, thus the tendency is to increase the number of simultaneous collisions. This make harder to reconstruct the events from the measures. The convolutive neural networks can solve similar problems but they are not adapted to the geometry of the detectors. This is why new techniques arise, namely graph convolutions, adapted to 3D data and leading to particle identification. The problem of these techniques is that they require graph constructions which mean complexity is quadratic and reduce their utilisability. But, as the sensors of the detectors have fixed geometry, it is possible to pre-compute the connectivity and to reduce the mean complexity of their construction.

The goal of this PhD is the develop different algorithms of graph construction and convolution applied on particle identification benchmarks: calorimeter and neutrino detector event classification, event segmentation (disentanglement), optimized and quantified version for electronics implementation and parallelized version. This PhD position is opened to any student with a good level in machine learning techniques. No physics prerequisites are necessary. Thus, students in computer science are welcome.

All the details are in the detailed proposition.