Robustness of the smartpixels classifier for different simulated sensor geometries and non-ideal detector conditions

Thu, 01 Jan 2026 00:00:00 +0000

This paper studies the co-design of planar pixel sensors and on-sensor neural networks for real-time data reduction in future high-granularity pixel tracking detectors.

Key points

Neural networks embedded in the readout chip can reject low-\(p_T\) tracks based on the geometric shape of charge clusters, enabling higher granularity and possible use of pixel data in the first-level trigger.
The authors systematically explore how NN efficiency and data-reduction performance depend on:
- Sensor geometry (pitch and depth)
- External magnetic field (Lorentz drift)
- Radiation damage
- Noise
Smaller pitch in the bending direction improves \(p_T\) discrimination; larger pitch can be partially compensated by greater sensor depth.
A magnetic field parallel to the sensor plane broadens clusters via Lorentz drift and improves network performance; its absence degrades background rejection by \(\mathcal{O}(10\%)\).
Radiation damage alters cluster shapes and reduces signal efficiency by \(\sim 30\)–\(60\%\), but nearly full performance is recovered by retraining the network.
Noise effects are largely mitigated by retraining on noise-injected data (performance remains within \(6\%\) of the noiseless baseline).

The work provides essential guidance for designing viable “smart pixel” detectors for upcoming collider experiments.

Edge Computing |

Robustness of the smartpixels classifier for different simulated sensor geometries and non-ideal detector conditions

Key points