Smart pixel sensors: implementing deep learning for ultrafast data reduction in next-generation particle detectors

The Large Hadron Collider (LHC) is a particle accelerator that produces the highest energy particle collisions ever made in a lab. Studying these collisions enables the discovery of new fundamental particles and provides more information about the characteristics of well-known particles. As a charged particle traverses through a pixel detector, it deposits charge on pixel sensors, leaving charge clusters. Tracked particles generally have transverse momenta pT > 0.2 GeV, allowing them to hit enough pixel sensor layers for their tracks to be reconstructed. Particles that do not generate enough hits for track reconstruction are said to be untracked. Future upgrades to the LHC will require smaller, more granular pixels which will generate petabytes of data per second. We will apply ultrafast neural networks to reject data from background on detector in real time. One method to reduce the amount of data being stored is to implement a neural network on an integrated circuit to read out shapes of charge clusters and classify these clusters as originating from high or low pT tracks. Current research has focused on training a neural network on tracked particles, but we will shift to rejecting untracked particle clusters. Preliminary results indicate that modest selections on these pixel clusters could enable more than an order of magnitude of data reduction, opening up the potential to access precision tracking information at 40 MHz for the first time.