Machine learning methods for subpixel trajectory reconstruction in discretized position detectors
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Machine learning methods for subpixel trajectory reconstruction in discretized position detectors".
Mira: Here's the gist of the paper. Machine learning methods recover sub-pixel muon position in a segmented scintillator array The transformer and CNN cut median position error 5–8× versus the energy-weighted centroid Muon tomography uses penetrating cosmic-ray muons to image dense materials and large structures.
Kai: Muon tomography reconstructs particle trajectories from hit positions measured in multiple detector planes.
Mira: In a pixelated scintillator detector, the signal from a single muon can be distributed across several cells, limiting the spatial resolution obtained from the discrete detector geometry.
Kai: Because the array is coarsely segmented, an inclined muon and its secondary particles can deposit energy in more than one cell, so the measured multi-cell signal pattern must be used to estimate the underlying particle position.
Paper summary: Kai: So to summarize this paper, they are evaluating machine learning methods for subpixel position and angular reconstruction in segmented particle detectors using cosmic ray muon events simulated in Geant4 >
Mira: The thesis is that a transformer neural network can improve position and trajectory reconstruction directly from the signal map of a coarse, segmented scintillator array >
Kai: They propose this transformer architecture separates the task into two parts: first, coarse cell classification and second, subpixel offset regression >
Mira: This is different from previous work because they are testing if these learned models can recover subpixel crossing positions directly from the signal pattern of a coarse array rather than just using charge centroid reconstruction >
Lev: That means they are bypassing some of the traditional physics-based assumptions about how we calculate that center point, which is interesting for pushing detection limits >
Kai: The comparison they run involves four approaches: a transformer neural network, a convolutional neural network, a multilayer perceptron, and an energy-weighted centroid method >
Mira: They found that the transformer and CNN produce median position errors of zero point one eight centimeters and zero point two three centimeters corresponding to approximately three percent of the six point two five centimeter cell width >
Lev: So they are demonstrating that these models can achieve resolutions that are smaller than the physical cell size itself, which is what subpixel resolution means >
Kai: The paper highlights that relative to the centroid method, these transformer and CNN methods result in six to eightfold reductions in median position error and five to sevenfold reductions in median angular error >
Mira: That suggests a substantial gain over traditional methods for image reconstruction when dealing with coarse detectors like this one >
Lev: It matters because if we can reliably get subpixel information from these signals, it fundamentally changes how much detail we can extract from the cosmic ray track data >
Conclusion: Kai: So, looking at the title, "Machine learning methods for subpixel trajectory reconstruction in discretized position detectors," it really puts it all together >
Mira: It’s about using machine learning to solve a specific problem: getting better position accuracy when your detector is segmented and coarse >
Lev: What this means practically is that for things like muon tomography, where you’re looking at large structures, having better position information directly translates to a clearer image of the underlying material >
Kai: The authors are showing that learned models can recover those small positional details without changing the actual physical detector geometry >
Mira: The main implication is that we can use these techniques to achieve finer spatial resolution in applications where traditional methods struggle with coarse segmentation and ambiguity >
Lev: But they also flag a limitation, which is that their improvement in error is clearest only when looking at the median and lower percentiles of events >
Kai: And there’s this caveat about those events with competing secondary electron signals causing a heavy error tail, particularly on the second selected detector plane >
Mira: So the paper suggests that while machine learning offers a way to get subpixel recovery, we still have to worry about those specific tricky event types if we want reliable performance in practice >
University of Wisconsin–Madison · Lucid Motors
physics.ins-det
Submitted: 2025-12-12
Updated: 2026-11-11
Comments: Code and data: https://doi.org/10.5281/zenodo.24143582
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 85/100
The gist: Machine learning methods recover sub-pixel muon position in a segmented scintillator array The transformer and CNN cut median position error 5–8× versus the energy-weighted centroid Introduction…
Key concepts
- Muon Tomography
- This technique uses cosmic-ray muons as particles that penetrate dense materials. By measuring where these muons hit detectors in multiple planes, researchers can reconstruct the path of the muon and image internal structures within those materials.
- Sub-pixel Position Recovery
- This refers to recovering position accuracy finer than the physical size of a single detector cell. Traditional methods struggle with this because a single particle's signal spreads across several cells, making it difficult to pinpoint its exact location.
- Transformer and CNN
- These are advanced machine learning models used for reconstruction. The transformer and convolutional neural network (CNN) proved superior to simpler models like the MLP or centroid method, demonstrating their ability to learn complex spatial patterns from detector signals to achieve high-resolution position estimates.
Terminology
Summary
Machine learning methods recover sub-pixel muon position in a segmented scintillator array The transformer and CNN cut median position error 5–8× versus the energy-weighted centroid
Introduction to Muon Tomography
Muon tomography uses penetrating cosmic-ray muons to image dense materials and large structures. Muon tomography reconstructs particle trajectories from hit positions measured in multiple detector planes. p. 1 Further improvements in spatial and angular
resolution would extend the range of structures that can be imaged and reduce the exposure required to form a useful
tomographic reconstruction. Muon tomography reconstructs particle trajectories from hit positions measured in multiple detector planes. In
a pixelated scintillator detector, however, the signal from a single muon can be distributed across several cells. The
resulting light pattern does not directly specify the muon’s crossing point and therefore limits the spatial resolution
obtained from the discrete detector geometry. As a charged particle traverses a scintillator, it excites the material along
its path, and the subsequent de-excitation produces visible light [6, 7]. Because the array is coarsely segmented, an
inclined muon and its secondary particles can deposit energy in more than one cell, so the measured multi-cell signal
pattern must be used to estimate the underlying particle position.Where the paper says it
Where the paper says it
Where the paper says it
Comparison of Reconstruction Approaches
The study compares four reconstruction approaches: a transformer neural network, a convolutional neural network (CNN), a multilayer perceptron (MLP), and an energy-weighted centroid. The transformer and CNN produce median position errors of 0. p. 1 Using cosmic-ray muon events simulated in Geant4 for an 8×8
scintillator array, we compare four reconstruction approaches: a transformer neural network, a
convolutional neural network (CNN), a multilayer perceptron (MLP), and an energy-weighted
centroid. The transformer and CNN produce median position errors of 0. The transformer and CNN produce median position errors of 0. 23 cm, respectively, corresponding to approximately 3% of the 6. • The transformer median position error is ∼3% of the 6. 25 cm cell width. 21◦
and 0. Relative to the centroid method, these correspond to six- to eightfold reductions
in median position error and five- to sevenfold reductions in median angular error.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Training Data and Simulation Details
The simulation uses cosmic-ray muon events simulated in Geant4 for an 8×8 scintillator array. Each detector plane is a 50 × 50 cm2 polyvinyltoluene (PVT) scintillator segmented into an 8 × 8 array of 64 cells. p. 2 1 [18, 19] using the FTFP_BERT hadronic physics list together
with Geant4 optical physics. Each detector plane is a 50 × 50 cm2 polyvinyltoluene (PVT) scintillator segmented into
an 8 × 8 array of 64 cells. Each cell measures 6. The source generates negative muons with kinetic energies sampled from a Gaussian distribution centered at 4 GeV
with 𝜎 = 2 GeV. Each cell is treated as an isolated,
fully absorbing counter: its surfaces have zero reflectivity, so a photon is recorded in the cell in which it was produced
and transported. Inter-cell optical crosstalk and photodetector light-sharing are therefore not modeled, and multicell signal patterns arise only from the muon and its secondary particles depositing energy in more than one cell.Where the paper says it
Where the paper says it
Where the paper says it
Performance of Learned Models
The transformer reaches 0. p. 4 The methods separate more clearly in median absolute error. The transformer reaches 0. • A hybrid transformer splits reconstruction into cell classification and offset regression. • Transformer and CNN cut median position and angular error 5–8× versus the energy-weighted centroid. • A hybrid transformer splits reconstruction into cell classification and offset regression. • Transformer and CNN cut median position and angular error 5–8× versus the energy-weighted centroid. The CNN has the lowest position
RMSE, while the transformer has the lowest mean and median position errors. The MLP has an RMSE similar to those
of the CNN and transformer but a larger median error. • A hybrid transformer splits reconstruction into cell classification and offset regression. • Transformer and CNN cut median position and angular error 5–8× versus the energy-weighted centroid. 25 cm cell width. These results show that
all three learned models recover sub-cell position information, with the transformer producing the lowest typical-event
error.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Error Analysis and Limitations
The improvement is clearest in the median and lower percentiles, while a small population of events with competing secondary-electron signals produces a heavy error tail, particularly on the second selected detector plane. The plane-2 RMSE values for the neural methods are 1. p. 5 This difference indicates
that the degradation is concentrated in a small number of events rather than distributed uniformly across the sample. The plane-2 RMSE values for the neural methods are 1. The plane-2 RMSE values for the neural methods are 1. 36 cm, approximately four to five times their plane-1
values of 0. Plane 2 has similar median errors but substantially larger RMSE for the learned methods. This difference indicates
that the degradation is concentrated in a small number of events rather than distributed uniformly across the sample. The broader centroid distribution reflects the propagation of position errors from both
detector planes into the reconstructed angle. The total position-error histogram (figure 8) shows median errors of 0. Highlights
Machine Learning Methods for Sub-Pixel Trajectory Reconstruction in Discretized Position
Detectors
Matthew Mark Romano, Zhengzhi Liu, JungHyun Bae
• Machine learning recovers sub-pixel muon position in a segmented scintillator array. • A hybrid transformer splits reconstruction into cell classification and offset regression. • A hybrid transformer splits reconstruction into cell classification and offset regression. • Transformer and CNN cut median position and angular error 5–8× versus the energy-weighted centroid. Machine Learning Methods for Sub-Pixel Trajectory
Reconstruction in Discretized Position Detectors
Matthew Mark Romanoa,∗
, Zhengzhi Liub
and JungHyun Baea
aDepartment of Nuclear Engineering and Engineering Physics, University of Wisconsin–Madison, , Madison, 53706, WI, USA
bLucid Motors, 7373 Gateway Blvd, Newark, 94560, CA, USA
A R T I C L E I N F O
Keywords:
Particle tracking detectors
Muon tomography
Neural networks
Transformer architectures
A B S T R A C T
We evaluate machine-learning methods for sub-pixel position and angular reconstruction in
segmented particle detectors. Using cosmic-ray muon events simulated in Geant4 for an 8×8
scintillator array, we compare four reconstruction approaches: a transformer neural network, a
convolutional neural network (CNN), a multilayer perceptron (MLP), and an energy-weighted
centroid. • A hybrid transformer splits reconstruction into cell classification and offset regression. • Transformer and CNN cut median position and angular error 5–8× versus the energy-weighted centroid.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion on Sub-Pixel Recovery
The results show that learned reconstruction can recover position information below the cell width without changing the simulated detector geometry. The improvement is clearest in the median and lower percentiles. p. 6 These results show that learned reconstruction can recover position information below the cell width without
changing the simulated detector geometry. The improvement is clearest in the median and lower percentiles. The improvement is clearest in the median and lower percentiles. A small
population of events with competing secondary-electron signals produces a heavy error tail, particularly on the second
selected detector plane, and reduces the apparent improvement in RMSE. Experimental validation should therefore include measured optical response, calibration variation, electronic noise,
thresholds, and plane-dependent effects. Future work should also test whether plane-specific training, additional
detector layers, or explicit treatment of secondary-electron events can reduce the large-error tail.Where the paper says it
Where the paper says it
Where the paper says it
Data Availability and Acknowledgements
The Geant4 simulation source code, the Python analysis and reconstruction pipeline, and the derived datasets supporting this article are openly available on Zenodo at https://doi.org/10.5281/zenodo. p. 7 ML for sub-pixel trajectory reconstruction in discretized detectors
Data availability
The Geant4 simulation source code, the Python analysis and reconstruction pipeline, and the derived datasets
supporting this article are openly available on Zenodo at https://doi. 5281/zenodo. 5281/zenodo. Acknowledgements
This work is supported by the College of Engineering and the Grainger Institute for Engineering at the University
of Wisconsin–Madison. Acknowledgements
This work is supported by the College of Engineering and the Grainger Institute for Engineering at the University
of Wisconsin–Madison. This research was supported in part by an appointment to the Oak Ridge National Laboratory
GRO Program, sponsored by the U. This research was supported in part by an appointment to the Oak Ridge National Laboratory
GRO Program, sponsored by the U. Department of Energy and administered by the Oak Ridge Institute for Science
and Education.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
References
[1] G. Bonomi et al. p. 7 References
[1] G. Bonomi et al. Bonomi et al. , Applications of cosmic-ray muons, Prog. , Applications of cosmic-ray muons, Prog. 112 (2020) 103768.Where the paper says it
Where the paper says it
Where the paper says it
[2] K.N. Borozdin et al. p. 7 112 (2020) 103768. Borozdin et al. Borozdin et al. , Radiographic imaging with cosmic-ray muons, Nature 422 (2003) 277.Where the paper says it
Where the paper says it
[3] J. Bae and S. Chatzidakis, The effect of cosmic ray muon momentum measurement for monitoring shielded special nuclear materials, arXiv:2109. p. 7 , Radiographic imaging with cosmic-ray muons, Nature 422 (2003) 277. Chatzidakis, The effect of cosmic ray muon momentum measurement for monitoring shielded special nuclear materials,
arXiv:2109. Chatzidakis, The effect of cosmic ray muon momentum measurement for monitoring shielded special nuclear materials,
arXiv:2109. 02470 (2022).Where the paper says it
Where the paper says it
[4] L.W. Alvarez et al. p. 7 02470 (2022). Alvarez et al. Alvarez et al. , Search for hidden chambers in the pyramids, Science 167 (1970) 832.Where the paper says it
Where the paper says it
[5] K. Morishima et al. p. 7 , Search for hidden chambers in the pyramids, Science 167 (1970) 832. Morishima et al. Morishima et al. , Discovery of a big void in Khufu’s Pyramid by observation of cosmic-ray muons, Nature 552 (2017) 386.Where the paper says it
Where the paper says it
[6] G.F. Knoll, Radiation Detection and Measurement, John Wiley & Sons (2000). p. 7 , Discovery of a big void in Khufu’s Pyramid by observation of cosmic-ray muons, Nature 552 (2017) 386. Knoll, Radiation Detection and Measurement, John Wiley & Sons (2000).Where the paper says it
[7] C. Grupen and B. p. 7 Knoll, Radiation Detection and Measurement, John Wiley & Sons (2000). Grupen and B. Grupen and B. Shwartz, Particle Detectors, 2nd ed. Shwartz, Particle Detectors, 2nd ed. , Cambridge University Press (2008), doi:10. , Cambridge University Press (2008), doi:10. 1017/CBO9780511534966.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[8] V. Radeka and R. p. 7 1017/CBO9780511534966. Radeka and R. Radeka and R. Boie, Centroid finding method for position-sensitive detectors, Nucl. Boie, Centroid finding method for position-sensitive detectors, Nucl. Methods 178 (1980) 543–554,
doi:10. Methods 178 (1980) 543–554,
doi:10. 1016/0029-554X(80)90836-8.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[9] G. Landi, Properties of the center of gravity as an algorithm for position measurements, Nucl. p. 7 1016/0029-554X(80)90836-8. Landi, Properties of the center of gravity as an algorithm for position measurements, Nucl. Landi, Properties of the center of gravity as an algorithm for position measurements, Nucl. Methods Phys. Landi, Properties of the center of gravity as an algorithm for position measurements, Nucl. Methods Phys. Landi, Properties of the center of gravity as an algorithm for position measurements, Nucl. Methods Phys. Landi, Properties of the center of gravity as an algorithm for position measurements, Nucl. Methods Phys. Methods Phys. A 485 (2002) 698.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[10] Y. LeCun, Y. Bengio and G. p. 7 A 485 (2002) 698. Bengio and G. Bengio and G. Hinton, Deep learning, Nature 521 (2015) 436.Where the paper says it
Where the paper says it
[11] A. Krizhevsky, I. p. 7 Hinton, Deep learning, Nature 521 (2015) 436. Krizhevsky, I. Krizhevsky, I. Sutskever and G. Sutskever and G. Hinton, ImageNet classification with deep convolutional neural networks, Adv. Hinton, ImageNet classification with deep convolutional neural networks, Adv. Neural Inf. Hinton, ImageNet classification with deep convolutional neural networks, Adv. Neural Inf. Neural Inf. 25 (2012) 1097.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[12] J.P. van Schayck et al. p. 7 25 (2012) 1097. van Schayck et al. van Schayck et al. , Sub-pixel electron detection using a convolutional neural network, Ultramicroscopy 218 (2020) 113091.Where the paper says it
Where the paper says it
[13] G. DeZoort et al. p. 7 , Sub-pixel electron detection using a convolutional neural network, Ultramicroscopy 218 (2020) 113091. DeZoort et al. DeZoort et al. , Charged particle tracking via edge-classifying interaction networks, Comput. , Charged particle tracking via edge-classifying interaction networks, Comput. 5 (2021) 26.Where the paper says it
Where the paper says it
Where the paper says it
[14] A. Vaswani et al. p. 7 5 (2021) 26. Vaswani et al. Vaswani et al. , Attention is all you need, Adv. Neural Inf. 30 (2017) 5998.Where the paper says it
Where the paper says it
Where the paper says it
[15] S. Caron et al. p. 7 30 (2017) 5998. Caron et al. Caron et al. , Trackformers: In search of transformer-based particle tracking for the high-luminosity LHC era, Eur. , Trackformers: In search of transformer-based particle tracking for the high-luminosity LHC era, Eur. C 85 (2025) 460,
doi:10. C 85 (2025) 460,
doi:10. 1140/epjc/s10052-025-14156-3.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[16] S. Van Stroud et al. p. 7 1140/epjc/s10052-025-14156-3. Van Stroud et al. Van Stroud et al. , Transformers for charged particle track reconstruction in high-energy physics, Phys. , Transformers for charged particle track reconstruction in high-energy physics, Phys. X 15 (2025) 041046,
doi:10. X 15 (2025) 041046,
doi:10. 1103/md46-yqgd.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[17] Z. Liang et al. p. 7 1103/md46-yqgd. Liang et al. Liang et al. , A muon scattering tomography system based on high spatial resolution scintillating detector, Nucl. Methods Phys. A 1091 (2026) 171790, doi:10.Where the paper says it
Where the paper says it
Where the paper says it
[18] S. Agostinelli et al. p. 7 A 1091 (2026) 171790, doi:10. Agostinelli et al. Agostinelli et al. , Geant4—a simulation toolkit, Nucl. Methods Phys. A 506 (2003) 250.Where the paper says it
Where the paper says it
Where the paper says it
[19] J. Allison et al. p. 7 A 506 (2003) 250. Allison et al. Allison et al. , Recent developments in Geant4, Nucl. Methods Phys. A 835 (2016) 186.Where the paper says it
Where the paper says it
Where the paper says it
[20] Particle Data Group, Review of Particle Physics, Phys. p. 7 A 835 (2016) 186. [20] Particle Data Group, Review of Particle Physics, Phys. [20] Particle Data Group, Review of Particle Physics, Phys. D 110 (2024) 030001.Where the paper says it
Where the paper says it
[21] D.P. Kingma and J. p. 7 D 110 (2024) 030001. Kingma and J. Kingma and J. Ba, Adam: A method for stochastic optimization, arXiv:1412.Where the paper says it
Where the paper says it
[22] I. Loshchilov and F. p. 7 Ba, Adam: A method for stochastic optimization, arXiv:1412. Loshchilov and F. Loshchilov and F. Hutter, Decoupled weight decay regularization, International Conference on Learning Representations (ICLR), 2019;
arXiv:1711.Where the paper says it
Where the paper says it
[M. M. Romano et al. p. 0 Highlights
Machine Learning Methods for Sub-Pixel Trajectory Reconstruction in Discretized Position
Detectors
Matthew Mark Romano, Zhengzhi Liu, JungHyun Bae
• Machine learning recovers sub-pixel muon position in a segmented scintillator array. • A hybrid transformer splits reconstruction into cell classification and offset regression. Highlights
Machine Learning Methods for Sub-Pixel Trajectory Reconstruction in Discretized Position
Detectors
Matthew Mark Romano, Zhengzhi Liu, JungHyun Bae
• Machine learning recovers sub-pixel muon position in a segmented scintillator array. • A hybrid transformer splits reconstruction into cell classification and offset regression. Highlights
Machine Learning Methods for Sub-Pixel Trajectory Reconstruction in Discretized Position
Detectors
Matthew Mark Romano, Zhengzhi Liu, JungHyun Bae
• Machine learning recovers sub-pixel muon position in a segmented scintillator array. • A hybrid transformer splits reconstruction into cell classification and offset regression. Highlights
Machine Learning Methods for Sub-Pixel Trajectory Reconstruction in Discretized Position
Detectors
Matthew Mark Romano, Zhengzhi Liu, JungHyun Bae
• Machine learning recovers sub-pixel muon position in a segmented scintillator array. • A hybrid transformer splits reconstruction into cell classification and offset regression. Highlights
Machine Learning Methods for Sub-Pixel Trajectory Reconstruction in Discretized Position
Detectors
Matthew Mark Romano, Zhengzhi Liu, JungHyun Bae
• Machine learning recovers sub-pixel muon position in a segmented scintillator array. • A hybrid transformer splits reconstruction into cell classification and offset regression. Highlights
Machine Learning Methods for Sub-Pixel Trajectory Reconstruction in Discretized Position
Detectors
Matthew Mark Romano, Zhengzhi Liu, JungHyun Bae
• Machine learning recovers sub-pixel muon position in a segmented scintillator array. • A hybrid transformer splits reconstruction into cell classification and offset regression. Highlights
Machine Learning Methods for Sub-Pixel Trajectory Reconstruction in Discretized Position
Detectors
Matthew Mark Romano, Zhengzhi Liu, JungHyun Bae
• Machine learning recovers sub-pixel muon position in a segmented scintillator array. • A hybrid transformer splits reconstruction into cell classification and offset regression. Highlights
Machine Learning Methods for Sub-Pixel Trajectory Reconstruction in Discretized Position
Detectors
Matthew Mark Romano, Zhengzhi Liu, JungHyun Bae
• Machine learning recovers sub-pixel muon position in a segmented scintillator array. • A hybrid transformer splits reconstruction into cell classification and offset regression.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[M. M. Romano et al.: Preprint submitted to Elsevier Page 7 of 13]
[M. M. Romano et al.: Preprint submitted to Elsevier Page 8 of 13]
[M. M. Romano et al.: Preprint submitted to Elsevier Page 9 of 13]
[M. M. Romano et al.: Preprint submitted to Elsevier Page 10 of 13]
[M. M. Romano et al.: Preprint submitted to Elsevier Page 11 of 13]
[M. M. Romano et al.: Preprint submitted to Elsevier Page 12 of 13]
[M. M. Romano et al.: Preprint submitted to Elsevier Page 13 of 13] p. 1 Romano et al. : Preprint submitted to Elsevier Page 1 of 13Where the paper says it
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
-
Improved AI system can recover sub-pixel position information directly from signal maps of coarse, segmented scintillator arrays by using a hybrid transformer architecture that splits reconstruction into cell classification and offset regression.
-
The improved system can achieve median position errors of 0.18 cm for the transformer, which corresponds to approximately 2.9% of the 6.25 cm cell width, significantly outperforming the energy-weighted centroid's median error of 1.43 cm on plane 2 alone.
-
The system can improve trajectory reconstruction by utilizing a transformer's self-attention mechanism to incorporate nonlocal correlations not covered by the receptive field of two-layer CNNs, leading to lower angular errors compared to other methods.
Abstract
We evaluate machine-learning methods for sub-pixel position and trajectory-angle reconstruction in scintillator detectors. Geant4 simulations use two 50 times 50 cm plastic-scintillator planes separated by 30 cm, each represented by an 8 times 8 grid of 64 touching computational cells of pitch 6.25 cm. The plates are optically continuous. Photons reaching any outer face are absorbed and scored once in the pre-step cell. We sum the collected photon energies in each cell and normalize the signals to the plane total. Photodetectors and electronics are not simulated. We compare a transformer, a convolutional neural network (CNN), a multilayer perceptron (MLP), and an energy-weighted centroid. We compute directions from the two reconstructed positions. On a common 19,053-event test set, transformer and CNN median position errors (per-event means of the two planes' radial errors) are 0.18 and 0.23 cm, or 2.9% and 3.6% of the cell width. Their median space-angle errors are 0.35 and 0.52, compared with 1.03 cm and 1.90 for the centroid and 4.21 for an idealized Monte Carlo reference retaining only the traversed cell. The transformer has the lowest median errors, whereas the CNN has the lowest position- and angular-RMSE point estimates. Large-error tails limit the RMSE gains. Learned methods recover sub-pixel information in this idealized response. Establishing gains for muon tomography requires realistic optical coupling, photodetector response, and electronics.
Sources
- The Effect of Cosmic Ray Muon Momentum Measurement for Monitoring Shielded Special Nuclear Materials
- Adam: A Method for Stochastic Optimization
- Decoupled Weight Decay Regularization
Related papers
- Cryogenic Mechanical Loss of GaAs/AlGaAs Crystalline Coatings
- The physical design of the superfluid helium-4 UCN source at CSNS
- A new diffuse reflector filament for additive manufacturing of 3D printing finely-segmented plastic scintillator
- Characterization of CRYO ASIC for charge readout in the nEXO experiment
- Reconstructing High-Fidelity Light Yield Maps for Surface LArTPCs Using Crossing Cosmic Muons
- OASIS-ERIS: Open Acquisition System for IEPE Sensors - Electrically Refined and Inter-device Synchronization