Speaker
Description
Patterns of charge deposited in pixelated particle tracking detectors are rich with kinematic information that is lost to ever increasing bandwidth constraints in high-luminosity environments. On-sensor inference of track geometry at the collision rate will grant next-generation triggering algorithms access to regions closer to the interaction point than ever before. We evaluate the fidelity of on-pixel track parameter regression algorithms in the presence of simulated thermal detector noise. Optimizing the choice charge thresholds of the detector’s 2-bit ADC during network training improves performance over a range of injected noise values. Without noise, the network selects lower thresholds to expose faint charge patterns that encode the track’s incident angle. We propose a denoising autoencoder based training methodology that enriches some network layers with tracking information lost below the noise floor without increasing architecture size. This framework reinforces noise-robustness to improve inference at the extreme sensor-edge.