Machine Learning-Based Reconstruction for Resistive Silicon Sensors

arXiv:2607.11585 · hep-ex, cs.LG, nucl-ex · Submitted 2026-07-13 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.

Kai: We're looking at the paper: "Machine Learning-Based Reconstruction for Resistive Silicon Sensors".

Mira: Here's the gist of the paper. Low-Gain Avalanche Diodes (LGADs) and AC-coupled Low-Gain Avalanche Diodes (AC-LGADs) are promising technologies for precision timing and four-dimensional tracking In this work, we study machine learning-based reconstruction and compression for resistive silicon sensors.

Kai: Sensor Technology Context Low-Gain Avalanche Diodes (LGADs) and AC-coupled Low-Gain Avalanche Diodes (AC-LGADs) are silicon sensors with internal gain that provide precision timing.

Mira: In the AC-LGAD variant

13, 14: , the AC pad is coupled to the resistive n+ layer through a dielectric layer.

Kai: This structure provides a 100% fill factor and enables good spatial resolution with a relaxed readout pitch.

Paper summary: Kai: So we're looking at this paper now called "Machine Learning-Based Reconstruction for Resistive Silicon Sensors". It’s about using machine learning to reconstruct where a particle hit inside these special silicon sensors, specifically low-gain avalanche diodes and their AC-coupled versions.

Mira: The core idea seems to be tackling the problem of charge sharing in resistive silicon sensors. They point out that when a particle hits, the signal doesn't just go to one pad; it spreads across multiple pads, which makes getting precise position information really tricky because you can get close to the electronic noise threshold <ref:2607.11585#pg2>.

Lev: From an engineering side, that charge spreading is a big hurdle for any reconstruction algorithm. It complicates things because the useful information ends up near where the noise floor is, and traditional matrix inversion methods can just break down at that point <ref:2607.11585#pg1>.

Kai: So what's their main claim here? They are looking at how much useful information you can actually recover from that noisy signal near the noise threshold for position reconstruction <ref:2607.11585#pg1>.

Mira: The paper introduces a couple of machine learning approaches to see if they can improve this recovery, starting with recurrent neural network models based on LSTM layers for full-waveform reconstruction <ref:2607.11585#pg3>. They also explore ways to reduce the data size first by doing waveform rasterization and window selection methods <ref:2607.11585#pg3>.

Lev: The paper mentions that these models are designed to work with arbitrary pad counts and geometries, which is good for flexibility, but they have to manage edge distortions too <ref:2607.11585#pg3>. They claim these approaches can preserve a position resolution of about ten micrometers for sensors that have a five hundred micrometer by five hundred micrometer pitch <ref:2607.11585#pg3>.

Kai: So, when they test the full-waveform reconstruction model, what did they find in terms of performance improvement over existing methods?

Mira: They found that the first full-waveform reconstruction model using the LSTM recurrent neural network achieved a fifty percent improvement over matrix inversion methods and a seven fold improvement over binary readout <ref:2607.11585#pg3>.

Lev: Seven fold is significant when you think about running this on actual hardware, because that implies a much smaller computational overhead compared to those older methods, even if the underlying model is complex <ref:2607.11585#pg3>.

Kai: They then moved to a newer topology agnostic transformer-based direction for reconstruction, and what did that architecture bring to the table?

Mira: The transformer based reconstruction reduces dependence on the specific geometry of the sensor, which helps mitigate those edge distortions they mentioned earlier <ref:2607.11585#pg3>. It also supports arbitrary pad counts and geometries <ref:2607.11585#pg3>.

Lev: That sounds promising for real hardware deployment because if you can make the reconstruction less sensitive to the exact physical layout, it simplifies the required calibration steps for your error correction scheme <ref:2607.11585#pg3>. They project a residual width of about ten micrometers and a median absolute residual of about six point six micrometers <ref:2607.11585#pg3>.

Kai: So, to put it simply for someone listening who isn't in this field, what is the main contribution of this paper? What’s the big picture here?

Mira: The main contribution seems to be combining a pitch normalized performance with a better way to model those rare events that happen when particles interact with these sensors <ref:2607.11585#pg3>. They achieved that ten micrometer projected residual width and six point six micrometer median absolute residual while using an architecture that doesn't depend on the sensor shape <ref:2607.11585#pg3>.

Lev: It’s interesting because the authors mention they are targeting application-specific designs by tuning things like sensor size, gain characteristics, and depletion layer size <ref:2607.11585#pg2>. That means the performance isn't just about the algorithm; it depends entirely on how you design that physical sensor to start with <ref:2607.11585#pg2>.

Kai: So, what does this mean for people working on timing detectors in general? What’s the practical implication of them focusing on these resistive silicon sensors with machine learning <ref:2607.11585#pg3>?

Mira: It suggests that machine learning is a viable path to getting high precision timing and four dimensional tracking, especially when dealing with the inherent charge sharing in these AC-LGADs <ref:2607.11585#pg2>. It shows that you can get good position reconstruction even with the signal being spread out across multiple pads <ref:2607.11585#pg2>.

Lev: For real hardware, it means if you can model the readout and sensor design together, you can achieve better timing performance by using machine learning to clean up that complex signal sharing before you try to run it through your error correction software <ref:2607.11585#pg3>.

Kai: So, to wrap up this part of the discussion about "Machine Learning-Based Reconstruction for Resistive Silicon Sensors," the paper is showing how ML can handle the complexity introduced by charge sharing in these AC-LGADs and achieve a very good spatial resolution of about ten micrometers <ref:2607.11585#pg3>.

Mira: And it points toward application-specific sensor designs where you tune the gain and geometry to get the best performance for your specific tracking needs <ref:2607.11585#pg2>.

Lev: It’s a solid foundation for how we think about putting these models onto actual hardware, showing that the computational overhead can be managed if you use architectures that are less sensitive to the exact sensor layout <ref:2607.11585#pg3>.

Conclusion: Kai: So we’re wrapping up this look at "Machine Learning-Based Reconstruction for Resistive Silicon Sensors." Basically, this paper is about using machine learning to figure out where particles hit in these resistive silicon sensors, which have a known problem called charge sharing.

Mira: Right. They took that charge sharing issue and tried to solve it with deep learning models, specifically recurrent neural networks and then transformers. The authors are showing how these models can give you a much better position guess than older methods, like matrix inversion or just looking at the raw binary signal.

Lev: From an engineering standpoint, what they’re really tackling is that noise threshold problem. They want to know how much useful information we can actually pull out of a signal that's sitting right near the electronic floor before it becomes pure static.

Kai: I think the authors are pushing for a topology agnostic reconstruction method using those transformer models. That means the algorithm should work whether you have a simple strip or a super complex, curved sensor geometry. It’s about making it less dependent on how exactly you built the hardware.

Mira: And they claim this new approach gets us to about ten micrometers in projected residual width and six point six micrometers median absolute residual for these five hundred by five hundred micrometer sensors. That's a specific number, but we need to know what that means for the physics application.

Lev: Ten micrometers is tight for tracking, but if we can get that with an architecture that doesn't care about the exact pad arrangement, that’s actually quite useful because it simplifies the calibration side of things. The overhead on real hardware is what I watch most closely here.

Kai: So, to sum up, this work isn't just about a new algorithm; it’s about showing that machine learning can handle the inherent messiness of signal sharing in these sensors and deliver a stable position estimate for high-precision tracking.

Mira: It suggests that for future detectors, designing the sensor size and gain characteristics to match what the machine learning model expects is going to be crucial for getting those good results.

Lev: Next up, we’re going to look at how this kind of reconstruction could actually run on real hardware with low power constraints.

Brown University · University of Zurich · Brookhaven National Laboratory · Northwestern University

hep-ex, cs.LG, nucl-ex

Submitted: 2026-07-13

Updated: 2026-11-09

Journal ref: 2026 JINST 21 C08015

DOI: 10.1088/1748-0221/21/08/C08015

Project page: https://sengerm.github.io/Chubut_2

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: Low-Gain Avalanche Diodes (LGADs) and AC-coupled Low-Gain Avalanche Diodes (AC-LGADs) are promising technologies for precision timing and four-dimensional tracking In this work…

Key concepts

Low-Gain Avalanche Diodes (LGADs)
These are silicon sensors designed for precision timing that have internal gain. They are used in particle detectors to measure the arrival time of particles with high accuracy. The AC-coupled variant uses a specific structure to improve spatial resolution and fill factor.
Charge Sharing Modeling
When a particle hits the sensor, the resulting electrical signal is not localized to a single readout pad; it spreads across multiple pads. This phenomenon, called charge sharing, is modeled using parameters like Si, which describes how the signal fraction relates to noise and gain.
Transformer-based Reconstruction
This is a newer machine learning architecture used for position reconstruction that does not depend on the specific physical layout (topology) of the sensor pads. It helps mitigate errors caused by edge distortions and supports various pad counts while maintaining high positional accuracy.

Terminology

Summary

Low-Gain Avalanche Diodes (LGADs) and AC-coupled Low-Gain Avalanche Diodes (AC-LGADs) are promising technologies for precision timing and four-dimensional tracking In this work, we study machine learning-based reconstruction and compression for resistive silicon sensors. p. 0

Sensor Technology Context

Low-Gain Avalanche Diodes (LGADs) and AC-coupled Low-Gain Avalanche Diodes (AC-LGADs) are silicon sensors with internal gain that provide precision timing. In the AC-LGAD variant [13, 14], the AC pad is coupled to the resistive n+ layer through a dielectric layer. p. 0 This structure provides a 100% fill factor and enables good spatial resolution with a relaxed readout pitch. p. 0 The sensor fabrication can be targeted to the needs of a specific application by tuning the sensor size, gain characteristics, and depletion-layer size. p. 1

Charge Sharing Modeling

Increased charge sharing is an intrinsic property of RSDs and AC-LGADs. When a particle interacts with the sensor, the induced signal is not confined to a single readout pad. p. 1 The signal fraction is parameterized as Si = P αi/ ln ri n i αi/ ln ri (1). p. 1 The main question this work is beginning to address is how much information near the noise threshold can be recovered and used for position reconstruction. p. 1

Machine Learning Approaches

We first introduce recurrent neural network models based on LSTM layers, which provide a proof-of-concept implementation for full-waveform reconstruction. We also study routes towards bandwidth reduction with waveform rasterisation and window-selection methods. p. 0 These models are designed to support arbitrary pad counts and geometries, mitigate edge distortions, preserve approximately 10 µm position resolution for 500 µm × 500 µm pitched sensors. p. 0

Reconstruction Performance

The first full-waveform reconstruction model uses a recurrent neural network with Long Short-Term Memory (LSTM) layers, and with this model, we achieve a 50% improvement over matrix-inversion methods and a seven-fold improvement over binary readout. The newer topology-agnostic transformer-based direction reduces geometry dependence, mitigates edge distortions, and supports arbitrary pad counts and geometries. p. 7 The transformer-based reconstruction yields a projected residual width of about 10 µm with a median absolute residual of about 6. p. 56 µm.

Conclusion and Future Work

The main contribution of the proposed approach pertains rather to the combination of competitive pitch-normalized performance with improved rare events modelling. The transformer-based reconstruction reaches a projected residual width of about 10 µm, with a median absolute residual of about 6. p. 56 µm while using a topology-agnostic architecture. p. 5 The model maintains a position resolution of around 10 µm for 500 µm × 500 µm pitched sensors. p. 5

References

[1] CMS Collaboration. A MIP timing detector for the CMS phase-2 upgrade, 2019

[2] CMS Collaboration. The CMS experiment at the CERN LHC. p. 7 Journal of Instrumentation, 3(08):S08004, 2008

[3] C. Royon and F. p. 7 p. 7 Gautier. Fast timing detectors with applications in cosmic ray physics and medical science. p. 7 p. 7 Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1058:168886, 2024

[4] F. Gautier et al. p. 7 p. 7 A novel technique for real-time ion identification and energy measurement for in situ space instrumentation. p. 7 p. 7 Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1012:165599, 2021

[5] X. Wu et al. Penetrating particle analyzer (PAN). p. 7 p. 7 Advances in Space Research, 63(8):2672–2682, 2019

[6] ECFA Detector R&D Roadmap Process Group. The 2021 ECFA detector research and development roadmap. p. 7 p. 7 CERN-ESU-017, 2020

[7] G. Pellegrini et al. p. 7 p. 7 Technology developments and first measurements of low gain avalanche detectors (LGAD) for high energy physics applications. p. 7 p. 7 Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 765:12–16, 2014

[8] K. Wu et al. Design and fabrication of low gain avalanche detectors (LGAD): a TCAD simulation study. p. 7 p. 7 Journal of Instrumentation, 15(03):C03008, 2020

[9] N. Cartiglia et al. p. 7 p. 7 LGAD designs for future particle trackers. p. 7 p. 7 Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 979:164383, 2020

[10] F. Siviero et al. p. 7 p. 7 p. 7 p. 7 Machine learning for precise hit position reconstruction in resistive silicon detectors. p. 7 p. 7 Journal of Instrumentation, 19(01):C01028, 2024

[11] G. Giacomini et al. p. 0 p. 0 p. 0 p. 0 Development of a technology for the fabrication of low-gain avalanche diodes at BNL. p. 7 p. 7 Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 934:52–57, 2019

[12] G. Giacomini et al. Fabrication of different LGAD-based devices at BNL. p. 7 p. 7 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC), pages 1–4, 2021

[13] M. Mandurrino et al. p. 7 p. 7 Analysis and numerical design of resistive AC-coupled silicon detectors (RSD) for 4D particle tracking. p. 8 p. 8 Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 959:163479, 2020

[14] M. Tornago et al. p. 7 p. 7 p. 7 p. 7 Resistive AC-coupled silicon detectors (RSD): Principles of operation and first characterization. p. 8 p. 8 p. 8 Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1003:165319, 2021

[15] A. Apresyan et al. p. 7 p. 7 p. 7 Measurements of an AC-LGAD strip sensor with a 120 GeV proton beam. p. 8 p. 8 p. 8 Journal of Instrumentation, 15(09):P09038, 2020

[16] R. Heller et al. Characterization of BNL and HPK AC-LGAD sensors with a 120 GeV proton beam. p. 8 p. 8 Journal of Instrumentation, 17(05):P05001, may 2022

[17] G. D’Amen et al. p. 7 p. 7 Measurements of time and spatial resolution of AC-LGADs with different designs. p. 8 p. 8 Journal of Instrumentation, 17(08):C08007, 2022

[18] F. Siviero et al. First application of machine learning algorithms to the position reconstruction in resistive silicon detectors. p. 8 p. 8 Journal of Instrumentation, 16(03):P03019, mar 2021

[19] M. Tornago et al. Silicon sensors with resistive readout: Machine learning techniques for ultimate spatial resolution. p. 8 p. 8 Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1047:167816, 2023

[20] G. Barone et al. p. 0 p. 0 From analog readout to ML-processed silicon device signal-sharing and LGADs at BNL, 2024

[21] A. Aoki et al. p. 0 p. 0 Studies on ML processing and compression of signal shared ACLGADs, 2025

[22] M. Senger. Chubut2 readout board. p. 2 p. 2 https://sengerm. p. 8 p. 8github.io/Chubut 2/

[23] FastML Team. fastmachinelearning/hls4ml, 2024

[24] F. Fahim et al. p. 8 p. 8 hls4ml: An open-source codesign workflow to empower scientific low-power machine learning devices, 2021

[25] J. Duarte et al. Fast inference of deep neural networks in FPGAs for particle physics. p. 8 Journal of Instrumentation, 13(07):P07027, jul 2018

[26] J-F. Schulte et al. p. 8 hls4ml: A flexible, open-source platform for deep learning acceleration on reconfigurable hardware. p. 8 2025

[27] R. Arcidiacono et al. p. 7 p. 7 High-precision 4D tracking with large pixels using thin resistive silicon detectors. p. 8 Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1057:168671, 2023

[28] R. Arcidiacono et al. Resistive AC-coupled silicon detectors (RSD): Principles of operation and first characterization. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1003:165319, 2021

[29] A. Apresyan et al. Measurements of an AC-LGAD strip sensor with a 120 GeV proton beam. Journal of Instrumentation, 15(09):P09038, 2020

[3] C. Royon and F. Gautier. Fast timing detectors with applications in cosmic ray physics and medical science. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1058:168886, 2024

[4] F. Gautier et al. A novel technique for real-time ion identification and energy measurement for in situ space instrumentation. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1012:165599, 2021

[5] X. Wu et al. Penetrating particle analyzer (PAN). Advances in Space Research, 63(8):2672–2682, 2019

[6] ECFA Detector R&D Roadmap Process Group. The 2021 ECFA detector research and development roadmap. CERN-ESU-017, 2020

[7] G. Pellegrini et al. Technology developments and first measurements of low gain avalanche detectors (LGAD) for high energy physics applications. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 765:12–16, 2014

[8] K. Wu et al. Design and fabrication of low gain avalanche detectors (LGAD): a TCAD simulation study. Journal of Instrumentation, 15(03):C03008, 2020

[9] N. Cartiglia et al. LGAD designs for future particle trackers. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 979:164383, 2020

[10] F. Siviero et al. Machine learning for precise hit position reconstruction in resistive silicon detectors. Journal of Instrumentation, 19(01):C01028, 2024

[11] G. Giacomini et al. Development of a technology for the fabrication of low-gain avalanche diodes at BNL. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 934:52–57, 2019

[12] G. Giacomini et al. Fabrication of different LGAD-based devices at BNL. 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC), pages 1–4, 2021

[13] M. Mandurrino et al. Analysis and numerical design of resistive AC-coupled silicon detectors (RSD) for 4D particle tracking. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 959:163479, 2020

[14] M. Tornago et al. Resistive AC-coupled silicon detectors (RSD): Principles of operation and first characterization. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1003:165319, 2021

[15] A. Apresyan et al. Measurements of an AC-LGAD strip sensor with a 120 GeV proton beam. Journal of Instrumentation, 15(09):P09038, 2020

[16] R. Heller et al. Characterization of BNL and HPK AC-LGAD sensors with a 120 GeV proton beam. Journal of Instrumentation, 17(05):P05001, may 2022

[17] G. D’Amen et al. Measurements of time and spatial resolution of AC-LGADs with different designs. Journal of Instrumentation, 17(08):C08007, 2022

[18] F. Siviero et al. First application of machine learning algorithms to the position reconstruction in resistive silicon detectors. Journal of Instrumentation, 16(03):P03019, mar 2021

[19] M. Tornago et al. Silicon sensors with resistive readout: Machine learning techniques for ultimate spatial resolution. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 1047:167816, 2023

[20] G. Barone et al. From analog readout to ML-processed silicon device signal-sharing and LGADs at BNL, 2024

[21] A. Aoki et al. Studies on ML processing and compression of signal shared ACLGADs, 2025

[22] M. Senger. Chubut2 readout board. https://sengerm.github.io/Chubut 2/

[23] FastML Team. fastmachinelearning/hls4ml, 2024

[24] F. Fahim et al. hls4ml: An open-source codesign workflow to empower scientific low-power machine learning devices, 2021 <ref:2607. p. 8

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Full-waveform reconstruction model using Long Short-Term Memory layers can predict hit position with a median Mean Squared Error of about 10 um for 500 um x 500 um pitched sensors, representing a five hundred percent improvement over matrix inversion methods and a seven fold improvement over binary readout.

  2. A topology agnostic transformer based model can be implemented to support arbitrary pad counts and geometries, mitigating edge distortions and yielding clusters closer to the true positions by using cross channel correlations that are part of the physical response.

  3. The transformer based reconstruction yields residual standard deviations of about 12 um in x, 11 um in y, and 6.6 um for the projected residual, which is substantially better than charge sharing and matrix inversion baselines when normalized to sensor pitch.

Related papers