Speaker
Description
Smartpixels explores on-sensor machine learning inference for silicon pixel detectors, targeting a future upgrade of the CMS detector in the High-Luminosity LHC era. By co-designing ML algorithms with readout ASIC constraints, charged-particle hit parameters are reconstructed directly on the sensor, cutting the data volume sent off-detector and potentially bringing pixel information within reach of the Level-1 trigger.
This talk presents training studies that incorporate the response of the front-end electronics, addressing a limitation of earlier Smartpixels studies. Clusters generated with PixelAV are propagated through a model of the readout front end, characterised from the circuit design, which applies the amplifier response, the correlated readout noise and the comparator time-walk, so that both the amplitude and the timing of the on-chip readout are represented. Networks are then trained in a sequence that removes one idealization at a time, from a Vision Transformer that learns the flash ADC thresholds jointly with its weights, to a compact convolutional network quantized at the bit widths of the target ASIC.