31 August 2026 to 4 September 2026
US/Pacific timezone
All in-person registration fee waivers have now been claimed.

A Comparative Study of Neural-Network Architectures for Embedded On-Chip Inference For Future MAPS Pixel Detectors

Not scheduled
12m
Presentation Contributed Talks

Speaker

Gian Michele Innocenti (Massachusetts Inst. of Technology (US))

Description

On-detector data reduction has become a primary constraint for next-generation monolithic active pixel sensors (MAPS) operated in continuous readout, which are used in particle physics to realize high-resolution, low-material-budget tracking and vertexing detectors. As pixel pitches decrease and hit rates increase, the bandwidth, power, and material-budget cost of this data transport is becoming a dominant constraint on detector design. On-chip inference provides a means to reduce the data volume at the source while preserving the information required for track reconstruction. The performance of this approach depends on the neural-network architecture and its implementation cost in silicon, which must be quantified through full hardware implementation. We present a systematic comparison of compact architectures for embedded on-chip inference, with the objective of characterizing the design space rather than optimizing a single configuration. Several architecture families are evaluated as a function of network size and quantization precision. For this exploratory study, each design point is implemented in a standard 28 nm CMOS process to obtain hardware-grounded metrics, including area, power, latency, and on-chip memory, and to determine how these scale with architectural and quantization choices. This process provides a cost-effective basis for comparing the embedded-AI architectures, capturing the relative trade-offs and scaling behavior that transfer to a MAPS-compatible CMOS imaging technology required for full sensor integration. The networks are trained on single-layer detector-response simulations calibrated against test-beam data and evaluated on a set of low-level tasks, including noise rejection, hit clustering, and the reconstruction of basic track observables. By comparing architectures on a common basis, this study quantifies their relative trade-offs and identifies the families compatible with MAPS-class tracking-layer constraints. The results provide guidance for selecting embedded-AI building blocks for future low-power, low-material-budget, AI-boosted MAPS sensors for next-generation tracking and vertexing detectors in particle and nuclear physics.

Authors

Abraham Thomas Holtermann (Massachusetts Inst. of Technology (US)) Gian Michele Innocenti (Massachusetts Inst. of Technology (US)) Ivan Amos Cali (Massachusetts Inst. of Technology (US)) Jelena Lalic (Massachusetts Inst. of Technology (US)) Marc DAVID Nichitiu (Massachusetts Inst. of Technology (US)) Pedro Vicente Leitao (Massachusetts Inst. of Technology (US)) Reng Zheng (MIT)

Presentation materials

There are no materials yet.