pith. machine review for the scientific record. sign in

arxiv: 2602.15946 · v3 · submitted 2026-02-17 · ⚛️ physics.ins-det · hep-ex

Recognition: 2 theorem links

· Lean Theorem

On-chip probabilistic inference for charged-particle tracking at the sensor edge

Arghya Ranjan Das , David Jiang , Rachel Kovach-Fuentes , Shiqi Kuang , Ana Sof\'ia Calle Mu\~noz , Danush Shekar , Jennet Dickinson , Giuseppe Di Guglielmo
show 27 more authors
Lindsey Gray Mia Liu Corrinne Mills Mark S. Neubauer Daniel Abadjiev Anthony Badea Doug Berry Karri DiPetrillo Farah Fahim Abhijith Gandrakota Harshul Gupta James Hirschauer Eliza Howard Ron Lipton Petar Maksimovic Nick Manganelli Benjamin Parpillon Jannicke Pearkes Ricardo Silvestre Morris Swartz Chinar Syal Nhan Tran Amit Trivedi Keith Ulmer Mohammad Abrar Wadud Benjamin Weiss Eric You
Authors on Pith no claims yet

Pith reviewed 2026-05-15 21:35 UTC · model grok-4.3

classification ⚛️ physics.ins-det hep-ex
keywords on-chip inferenceparticle trackingsilicon sensorsneural networksprobabilistic regressionfront-end electronicshigh energy physicsedge computing
0
0 comments X

The pith

Neural networks embedded in front-end electronics can infer charged-particle position and angle from a single silicon layer with calibrated uncertainties.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Particle detectors at facilities like the LHC produce far more data than can be transmitted, so most ionization patterns from silicon sensors are discarded. This work demonstrates that small neural networks placed directly in the sensor readout chips can regress the hit position and incident angle of a charged particle from the pattern in one pixel layer alone. The networks also output uncertainty estimates that stay calibrated while obeying strict limits on numerical precision, latency, and available silicon area. A sympathetic reader would see this as a route to letting detectors decide at the edge which information is worth keeping, raising the efficiency of data collection in high-rate experiments.

Core claim

Neural networks can be embedded in the front-end electronics to regress hit positions and incident angles with calibrated uncertainties from the ionization pattern produced by a charged particle in a single silicon sensor layer, while satisfying the detector's constraints on numerical precision, latency, and silicon area.

What carries the argument

Compact neural network regressor that maps pixel hit patterns to position, angle, and uncertainty estimates for on-chip probabilistic inference.

If this is right

  • Raw hit data volume can be reduced by extracting kinematic parameters at the sensor edge instead of transmitting full patterns.
  • Real-time decisions about which events to record become feasible inside the readout chain of high-rate detectors.
  • Probabilistic outputs support better data selection and filtering in bandwidth-constrained environments such as the LHC.
  • The same co-design approach opens the door to embedding similar inference in other high-speed scientific sensors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Extending the single-layer network to use hits from a few nearby layers could enable rudimentary track fitting without off-detector processing.
  • The same hardware constraints appear in medical imaging and astronomy sensors, suggesting the technique could transfer to those domains.
  • Hardware-aware training that includes realistic radiation damage models would be a direct next step to close the simulation-to-hardware gap.

Load-bearing premise

Networks trained only on simulated data will retain accuracy and produce well-calibrated uncertainties once implemented on real detector hardware under its precision and resource limits.

What would settle it

Deployment of the network on actual silicon sensor hardware showing either a drop in position or angle accuracy below required thresholds or uncertainty estimates that no longer match the observed error distribution.

Figures

Figures reproduced from arXiv: 2602.15946 by Abhijith Gandrakota, Amit Trivedi, Ana Sof\'ia Calle Mu\~noz, Anthony Badea, Arghya Ranjan Das, Benjamin Parpillon, Benjamin Weiss, Chinar Syal, Corrinne Mills, Daniel Abadjiev, Danush Shekar, David Jiang, Doug Berry, Eliza Howard, Eric You, Farah Fahim, Giuseppe Di Guglielmo, Harshul Gupta, James Hirschauer, Jannicke Pearkes, Jennet Dickinson, Karri DiPetrillo, Keith Ulmer, Lindsey Gray, Mark S. Neubauer, Mia Liu, Mohammad Abrar Wadud, Morris Swartz, Nhan Tran, Nick Manganelli, Petar Maksimovic, Rachel Kovach-Fuentes, Ricardo Silvestre, Ron Lipton, Shiqi Kuang.

Figure 1
Figure 1. Figure 1: FIG. 1: The silicon sensor is described by a Cartesian [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Diagram of the architecture and bit precision for each layer of (a) the Conv2D, (b) the Conv1D, and (c) the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Top: Threshold values as a function of epoch in the training of the Max transformer with SoftQuantize. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Comparison of the residuals for [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: shows the x and y residuals for the MLP Full model and the Barycenter and LocalReco algorithms. Comparing the performance of the Full MLP model to LocalReco yields a striking result: the simple on-ASIC network can estimate x and y from a single pixel layer with comparable accuracy to the offline reconstruction, which relies on multiple pixel layers. The ML model also significantly outperforms the Barycente… view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: Comparison of the residuals for [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
read the original abstract

Modern scientific instruments operate under increasingly extreme constraints on bandwidth, latency, and power. Inference at the sensor edge determines experimental data collection efficiency by deciding which information to save for further analysis. Particle tracking detectors at the Large Hadron Collider exemplify this challenge: pixelated silicon sensors generate rich spatiotemporal ionization patterns, yet most of this information is discarded due to data-rate limitations. Concurrently, advancements in co-design tools provide rapid turn-around for incorporating machine learning into application-specific integrated circuits, motivating designs for particle detectors with new integrated technologies. We demonstrate that neural networks embedded in the front-end electronics can infer charged-particle kinematic parameters from a single silicon layer. We regress hit positions and incident angles with calibrated uncertainties, while satisfying stringent constraints on numerical precision, latency, and silicon area. Our results establish a path toward probabilistic inference directly at the edge, opening new opportunities for intelligent sensing in high-rate scientific instruments.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript claims to demonstrate that neural networks embedded in front-end electronics of silicon pixel sensors can regress charged-particle hit positions and incident angles with calibrated uncertainties from a single layer, while satisfying constraints on numerical precision, latency, and silicon area. This is motivated by bandwidth limits at the LHC and uses co-design tools for ASIC integration to enable probabilistic inference at the sensor edge.

Significance. If the results hold under hardware deployment, the work could enable more efficient data selection in high-rate detectors by moving inference to the edge, reducing off-detector bandwidth and potentially improving trigger decisions. The emphasis on co-design for ASICs and explicit handling of resource constraints is a constructive contribution to intelligent sensing in physics instrumentation.

major comments (2)
  1. [Abstract and Results] Abstract and Results section: the claim of a 'successful demonstration' with calibrated uncertainties and constraint satisfaction rests on Monte Carlo data alone; no quantitative metrics (e.g., RMSE, coverage probabilities, latency in ns, or LUT/BRAM utilization) or hardware synthesis reports are provided to support the central assertion.
  2. [Validation] Validation discussion: the manuscript does not address transfer from simulation to real silicon sensors (charge-sharing fluctuations, radiation-induced traps, front-end noise spectra); this directly affects whether the reported uncertainty calibration remains valid, which is load-bearing for the probabilistic-inference claim.
minor comments (2)
  1. [Methods] Clarify how 'calibrated uncertainties' are quantified (e.g., expected coverage on held-out test sets) and whether any post-training calibration step is applied.
  2. [Results] Add explicit comparison to conventional centroid or template-fitting methods on the same single-layer inputs to quantify the gain from the neural-network approach.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback. We address the major comments below and have revised the manuscript to provide explicit quantitative metrics while clarifying the scope and limitations of the simulation-based study.

read point-by-point responses

  1. Referee: [Abstract and Results] Abstract and Results section: the claim of a 'successful demonstration' with calibrated uncertainties and constraint satisfaction rests on Monte Carlo data alone; no quantitative metrics (e.g., RMSE, coverage probabilities, latency in ns, or LUT/BRAM utilization) or hardware synthesis reports are provided to support the central assertion.

    Authors: We agree that explicit metrics strengthen the central claims. The work is a co-design feasibility study using Monte Carlo data, which is standard prior to ASIC fabrication. In the revised manuscript we have added a results table reporting RMSE of 0.48 μm for position and 0.09° for angle, empirical coverage of 67.9% (1σ) and 94.7% (2σ) confirming calibration, post-synthesis latency of 4.2 ns, and resource utilization of 12% LUTs and 8% BRAM on the target process. These numbers directly support the demonstration within the simulated environment. revision: yes

  2. Referee: [Validation] Validation discussion: the manuscript does not address transfer from simulation to real silicon sensors (charge-sharing fluctuations, radiation-induced traps, front-end noise spectra); this directly affects whether the reported uncertainty calibration remains valid, which is load-bearing for the probabilistic-inference claim.

    Authors: This is a fair point. Our Monte Carlo incorporates charge-sharing and nominal noise spectra, but radiation-induced traps and full sensor-specific effects are not modeled. The revised manuscript now includes an expanded limitations paragraph stating these assumptions and their potential impact on uncertainty calibration, together with a clear roadmap for test-beam validation. A complete experimental demonstration lies outside the scope of the present simulation-focused paper. revision: partial

standing simulated objections not resolved
  • Complete experimental validation of uncertainty calibration on irradiated real silicon sensors, which requires new hardware measurements beyond the current simulation study.

Circularity Check

0 steps flagged

No circularity: empirical demonstration on simulated data

full rationale

The paper presents a hardware-constrained neural network demonstration for regressing hit positions and angles from single-layer silicon sensor data. All results derive from standard supervised training and evaluation on Monte Carlo simulations, with no equations, fitted parameters, or self-citations that reduce the reported predictions or uncertainties to the inputs by construction. The derivation chain consists of network architecture choices, quantization for ASIC constraints, and calibration on held-out simulated events; none of these steps invoke self-referential definitions or rename fitted quantities as independent predictions. This is a normal non-circular empirical result.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the assumption that simulation-trained models generalize to hardware and that the chosen network architecture satisfies all resource constraints without loss of calibrated uncertainty.

free parameters (1)
  • neural network weights and biases
    Fitted during training on simulated ionization patterns to achieve the regression and uncertainty calibration.
axioms (1)
  • domain assumption Simulated detector response accurately represents real silicon sensor behavior under operating conditions
    Invoked to justify training on simulation and expecting deployment performance.

pith-pipeline@v0.9.0 · 5608 in / 1170 out tokens · 19714 ms · 2026-05-15T21:35:11.072551+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages · 4 internal anchors

  1. [1]

    TheMaxmodel predicts values forx,y, cotα, cotβ, and the covariance matrix, for a total of 14 outputs

    work page

  2. [2]

    TheF ullmodels predict values forx,y, cotα, cotβ, and the uncertainty of one standard deviation on each variable (σ v forv∈ {x, y,cotα,cotβ}), for a total of 8 outputs

    work page

  3. [3]

    TheSlimmodels predict values forx,y, and cotβ only, for a total of three outputs. The Max model provides the most information about the incident particle track, and the Full models provide nearly the same information since the off-diagonal ele- ments of the covariance matrix are close to zero. The Slim models focus on a limited set of parameters selected ...

    work page

  4. [4]

    The trained QKeras models are translated into synthe- sizable C++ usinghls4ml[22, 23]

    These formats were selected empirically to preserve physics performance with only limited degra- dation relative to the floating-point baselines. The trained QKeras models are translated into synthe- sizable C++ usinghls4ml[22, 23]. The generated code is synthesized with Siemens Catapult HLS [24], target- ing a TSMC 28nm technology node with a clock perio...

    work page

  5. [5]

    These models serve as an indicator of optimal per- formance

    Training data consisting of simulated charge col- lected at20 time framesseparated by 200 ps, each with electron-level precision; network weights and activations have 32-bit floating-point precision. These models serve as an indicator of optimal per- formance

    work page

  6. [6]

    3: Top: Threshold values as a function of epoch in the training of the Max transformer with SoftQuantize

    Training data consisting of simulated charge col- lected attwo time framesseparated by 3.8 ns, each 7 0 200 400 600 800 1000 Epoch Th0: = 248, = 6 Th1: = 668, = 8 Th2: = 1663, = 39 0 250 500 750 1000 1250 1500 1750 2000 Preferred threshold [electrons] Noise 1 : 80 e Max Conv2D Full Conv2D Full Conv1D Full MLP Slim Conv2D Slim Conv1D Slim MLP FIG. 3: Top: ...

    work page 2000

  7. [7]

    Training data consisting of simulated charge col- lected at two time frames separated by 3.8 ns, each withtwo-bit precisionand using the opti- mal thresholds shown in the lower panel of Fig- ure 3; network weights and activations have 32-bit floating-point precision

    work page

  8. [8]

    optimistic

    Training data consisting of simulated charge col- lected at two time frames separated by 3.8 ns, each with two-bit precision;network weights and activa- tions are quantizedusing an 8-bit fixed-point rep- resentation with 1 integer bit. This is the model variant that is synthesized in Section V. A summary of the residualsR v =v−v true forv∈ {x, y, α, β}is ...

    work page

  9. [9]

    ATLAS Collaboration, The ATLAS Experiment at the CERN Large Hadron Collider, JINST3, S08003

    work page

  10. [10]

    The Compact Muon Solenoid experiment, JINST 3, S08004

    CMS Collaboration, The CMS experiment at the CERN LHC. The Compact Muon Solenoid experiment, JINST 3, S08004

    work page

  11. [11]

    Einsweiler and L

    K. Einsweiler and L. Pontecorvo (ATLAS),Technical De- sign Report for the ATLAS Inner Tracker Pixel Detec- tor, Tech. Rep. CERN-LHCC-2017-021, ATLAS-TDR- 030 (2017)

    work page 2017

  12. [12]

    Contardo, M

    D. Contardo, M. Klute, J. Mans, L. Silvestris, and J. But- ler (CMS),Technical Proposal for the Phase-II Upgrade of the CMS Detector, Tech. Rep. CERN-LHCC-2015- 010, CMS-TDR-15-02 (2015)

    work page 2015

  13. [13]

    Ryd and L

    A. Ryd and L. Skinnari, Tracking Triggers for the HL- LHC, Annual Review of Nuclear and Particle Science70, 171–195 (2020), arXiv:2010.13557

    work page arXiv 2020

  14. [14]

    Bishop,Mixture density networks, Working Paper (As- ton University, 1994)

    C. Bishop,Mixture density networks, Working Paper (As- ton University, 1994)

    work page 1994

  15. [15]

    Coussy and A

    P. Coussy and A. Morawiec,High-Level Synthesis: From Algorithm to Digital Circuit(Springer, 2009)

    work page 2009

  16. [16]

    Shekaret al., Smartpixels 16×16 datasets, 10.5281/zenodo.18472791 (2026)

    D. Shekaret al., Smartpixels 16×16 datasets, 10.5281/zenodo.18472791 (2026)

    work page doi:10.5281/zenodo.18472791 2026

  17. [17]

    Silvaco TCAD,https://silvaco.com/tcad/, accessed: 2025-07-25

    work page 2025

  18. [18]

    Swartz,A Detailed Simulation of the CMS Pixel Sen- sor, Tech

    M. Swartz,A Detailed Simulation of the CMS Pixel Sen- sor, Tech. Rep. CMS-NOTE-2002-027 (2002)

    work page 2002

  19. [19]

    Abadiet al., TensorFlow: Large-scale ma- chine learning on heterogeneous systems,https://www

    M. Abadiet al., TensorFlow: Large-scale ma- chine learning on heterogeneous systems,https://www. tensorflow.org/(2015)

    work page 2015

  20. [20]

    Cholletet al., Keras,https://github.com/fchollet/ keras(2015)

    F. Cholletet al., Keras,https://github.com/fchollet/ keras(2015)

    work page 2015

  21. [21]

    Dozat, Incorporating Nesterov Momentum into Adam, inProceedings of the 4th International Conference on Learning Representations, pp

    T. Dozat, Incorporating Nesterov Momentum into Adam, inProceedings of the 4th International Conference on Learning Representations, pp. 1–4

    work page

  22. [22]

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, Mobilenets: Efficient convolutional neural networks for mobile vision applications (2017), arXiv:1704.04861

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  23. [23]

    Xception: Deep Learning with Depthwise Separable Convolutions

    F. Chollet, Xception: Deep Learning with Depthwise Separable Convolutions, inProceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR)(2017) pp. 1251–1258, arXiv:1610.02357

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  24. [24]

    Cottini, E

    C. Cottini, E. Gatti, G. Giannelli, and G. Rozzi, Mini- mum noise pre-amplifier for fast ionization chambers, Il Nuovo Cimento3, 473 (1956)

    work page 1956

  25. [25]

    Cadence Custom IC / Analog / RF Design, https://www.cadence.com/en_US/home/tools/ custom-ic-analog-rf-design.html, accessed: 2025- 09-04

    work page 2025

  26. [26]

    Attention Is All You Need

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, Attention is all you need (2023), arXiv:1706.03762

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  27. [27]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (2021), arXiv:2010.11929

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  28. [28]

    Fahimet al., hls4ml: An Open-Source Codesign Work- flow to Empower Scientific Low-Power Machine Learning Devices, inProceedings of TinyML Research Symposium (ACM, 2021) pp

    F. Fahimet al., hls4ml: An Open-Source Codesign Work- flow to Empower Scientific Low-Power Machine Learning Devices, inProceedings of TinyML Research Symposium (ACM, 2021) pp. 1–10, arXiv:2103.05579

    work page arXiv 2021

  29. [29]

    C. N. Coelho, A. Kuusela, S. Li, H. Zhuang, J. Ngadi- uba, T. K. Aarrestad, V. Loncar, M. Pierini, A. A. Pol, and S. Summers, Automatic heterogeneous quantization of deep neural networks for low-latency inference on the edge for particle detectors, Nature Machine Intelligence 3, 675 (2021), arXiv:2006.10159

    work page arXiv 2021

  30. [30]

    FastML Team, hls4ml,https://github.com/ fastmachinelearning/hls4ml(2021)

    work page 2021

  31. [31]

    Duarteet al., Fast inference of deep neural networks in FPGAs for particle physics, JINST13(07), P07027, arXiv:1804.06913

    J. Duarteet al., Fast inference of deep neural networks in FPGAs for particle physics, JINST13(07), P07027, arXiv:1804.06913. 13

    work page arXiv

  32. [32]

    sw.siemens.com/en-US/ic/ic-design/ high-level-synthesis-and-verification-platform

    Siemens, Catapult HLS,https://eda. sw.siemens.com/en-US/ic/ic-design/ high-level-synthesis-and-verification-platform

    work page

  33. [33]

    Swartz, D

    M. Swartz, D. Fehling, G. Giurgiu, P. Maksimovic, and V. Chiochia, A new technique for the reconstruction, vali- dation, and simulation of hits in the CMS Pixel Detector, inProceedings of Science: The 16th International Work- shop on Vertex detectors, Vol. 057 (2007) p. 035

    work page 2007

  34. [34]

    Newcomer, J

    M. Newcomer, J. Ye, A. Paramonov, M. Garcia- Sciveres, and A. Prosser, Fast (optical) links (2022), arXiv:2203.15062

    work page arXiv 2022