15–18 Dec 2026
Duke University
America/New_York timezone
Artificial Intelligence for the Electron Ion Collider (2026)

On-Chip AI for Data Compression in Future EIC Silicon Detectors

Not scheduled
20m
LSRC B101 Love Auditorium (Duke University)

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
AI/ML for ePIC and Beyond AI/ML for ePIC and Beyond

Speaker

Pedro Vicente Leitao (Massachusetts Institute of Technology)

Description

Precision tracking at the Electron-Ion Collider (EIC) requires highly granular silicon detectors with stringent constraints on power consumption and material budget. Transmitting raw detector data increases demands on readout bandwidth, power delivery, and cooling. On-chip AI offers a route to reducing data volume before transmission, aiming to preserve the information needed for tracking while potentially reducing the material budget of the associated detector infrastructure.

We present a framework for developing and benchmarking AI-based data compression in monolithic active pixel sensors (MAPS), with potential applications to future upgrades of ePIC or a second detector at the EIC [1–3]. This study draws on experience gained in the design and testing of 65 nm MAPS for the ePIC Silicon Vertex Tracker (SVT). Compact, quantized neural networks are trained using realistic single-layer sensor-response simulations derived from electron–proton events and constrained by test-beam measurements. Their performance is compared with conventional data-reduction algorithms using common detector inputs and a unified hardware implementation workflow.

The study connects model development and quantization to hardware synthesis and physical implementation, examining how architectural choices affect the balance between data reduction, retained tracking information, and implementation cost. We report results demonstrating substantial data reduction on simulated detector data, together with power, area, and latency estimates from digital implementation. These results guide the selection of compact AI models for future sensor integration, with the goal of enabling efficient on-chip compression for low-power, low-material-budget EIC tracking detectors.

[1] Abraham Thomas Holtermann, Gian Michele Innocenti, Ivan Amos Cali, Jelena Lalic, Marc David Nichitiu, Pedro Vicente Leitao and Reng Zheng, “A Comparative Study of Neural-Network Architectures for Embedded On-Chip Inference For Future MAPS Pixel Detectors.” Manuscript in preparation.
Related conference presentation: https://indico.cern.ch/event/1654479/contributions/7189072/

[2] Abraham Thomas Holtermann, Gian Michele Innocenti, Ivan Amos Cali, Jelena Lalic, Marc David Nichitiu, Pedro Vicente Leitao and Reng Zheng, “End-to-end benchmarking of on-detector data-reduction strategies: from realistic MAPS data to PPA.” Manuscript in preparation.
Related conference presentation: https://indico.cern.ch/event/1654479/contributions/7237463/

[3] Abraham Thomas Holtermann, Gian Michele Innocenti, Ivan Amos Cali, Jelena Lalic, Marc David Nichitiu, Pedro Vicente Leitao and Reng Zheng, “AI on ASICs in high-energy physics and a neural-network-based readout for future MAPS pixel detectors.” PIXEL2026, 11 September 2026.
https://indico.cern.ch/event/1519426/contributions/7162600/

Authors

Abraham Thomas Holtermann (Massachusetts Institute of Technology) Gian Michele Innocenti (Massachusetts Institute of Technology) Ivan Amos Cali (Massachusetts Institute of Technology) Jelena Lalic (Massachusetts Institute of Technology) Marc David Nichitiu (Massachusetts Institute of Technology) Pedro Vicente Leitao (Massachusetts Institute of Technology) Reng Zheng (Massachusetts Institute of Technology)

Presentation materials

There are no materials yet.