Search for the 16 O(ppp) to 13 C pi+ pi+ e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques
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: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Search for the 16 O(ppp) to 13 C pi+ pi+ e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques".
Mira: Here's the gist of the paper. Search for the 16O(ppp) →13 Cπ +π +e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques is a study that uses machine learning to search for nucleon decay, specifically focusing on the trinucleon decay mode in oxygen nuclei, which has implications for physics beyond the Standard Model The authors report a new partial lifetime limit of 4.2 × 1032 years for the trinucleon decay mode 16O(ppp) →13 Cπ +π +e+ obtained from a search conducted using the Super-Kamiokande detector with 0.401 megaton-years of exposure across five operational periods This represents an improvement of six orders of magnitude over previous experimental constraints Theoretical Motivation and Signal Definition The study focuses specifically on processes with ∆B = 3, which can occur through dimension-15 operators and represent the lowest-order baryon number violation permitted by this symmetry The specific process investigated is trinucleon decay in oxygen nuclei, where 16O(ppp) undergoes decay into 13Cπ+π+e+ Theoretical predictions for such processes estimate their partial lifetime to be on the order of 1033 years This channel is considered an especially sensitive probe of baryon-parity-protected scenarios because it respects B−L conservation and can be embedded in unified constructions that address neutrino masses and the origin of the baryon asymmetry Data Handling and Simulation Pipeline The analysis utilized Monte Carlo (MC) simulations for both the trinucleon decay signal and atmospheric neutrino background for each SK detector period The simulation of the trinucleon decay takes into account Fermi motion, correlations between nucleons, and nuclear binding energy Pion-nucleon interactions within the nucleus (pion FSI) are simulated using the NEUT cascade model to address final state interactions Atmospheric neutrinos constitute the primary background, with dominant processes including charged-current single pion production (CC1π) and deep inelastic scattering (DIS) Machine Learning Methodology The analysis employs a convolutional neural network (CNN) incorporating an attention mechanism to enhance event classification and improve sensitivity The CNN architecture is a modified version of the MobileNetV3 architecture, which uses Squeeze-and-Excitation mechanisms for adaptive weighting and a Bottleneck structure for dimensional reduction The detector response data is projected into a 2D plane using a Mercator projection to create an input array of dimensions
3, 224, 224: for the CNN Performance and Systematic Uncertainties The trained CNN model was applied to pre-selected samples from all five SK periods, producing a distribution with two distinct peaks separating the signal from the background The optimal cut value is determined independently for each SK data-taking period (SK-I through SK-V) to ensure consistent signal selection efficiency while maintaining effective background rejection Systematic uncertainties are categorized into physics modeling and detector/reconstruction components, with the FSI of pions being the dominant systematic uncertainty affecting signal efficiency at approximately 22–26% The final 90% confidence-level upper limit on the decay rate is established by combining posterior distributions from all five SK run periods The search resulted in only one event remaining in the signal region, which was consistent with an expected background of approximately 1.33 events, showing no statistically significant excess Consequently, a lower limit on the trinucleon decay lifetime at τ > 4.2 × 1032 years is established This result achieves more than a six-order-of-magnitude improvement over previous limits and reaches the theoretical model prediction for this ∆B = 3 mode References
1: A.
Kai: D.
Mira: Sakharov, “Violation of CP invariance, C asymmetry, and baryon asymmetry of the universe,” JETP Lett.
Kai: 5, 24-27 (1967)
2: H.
Title and authors: Kai: So, this paper is basically saying they used machine learning to scan massive amounts of data from Super-Kamiokande looking for nucleon decay, specifically this particular decay mode in oxygen nuclei into carbon, pions, and an electron.
Mira: Yeah, it’s about using that machine learning approach to hunt for something beyond the Standard Model through this trinucleon decay process. It’s important because if that happens, it means baryon number violation is happening in a way we don't expect one.
Lev: So what they actually report is a new partial lifetime limit of four point two times ten cubed years for this specific decay mode. That's an improvement over previous limits, six orders of magnitude better.
Kai: Six orders of magnitude is huge. The theoretical prediction for that process itself suggests a lifetime around one thousand thirty-three years three. So they found a limit that actually matches what theory expected.
Mira: That’s what makes it interesting, but we have to remember this is just a lower limit, not a detection. They are using Monte Carlo simulations for both the signal and the background across five different periods of Super-Kamiokande <ref:2609.25681#pg3>.
Lev: And that simulation part is where things get tricky because they have to model nuclear effects like pion final state interactions using a NEUT cascade model twenty. That’s a lot of physics modeling before the machine learning even gets involved.
Kai: Right, and the machine learning part uses a convolutional neural network with an attention mechanism to classify the events thirty-two. They used this modified MobileNetV3 architecture, which is pretty common but they adapted it for this specific detector response.
Mira: I’m also interested in how they handled the data projection. They use a cylindrical-to-planar projection mapping procedure, and they say that this transformation preserves topological properties of Cherenkov ring patterns three. That means the network learns things that are geometrically stable.
Lev: That geometric consistency is important because if the shape of those rings changes based on where you are in the detector, the network gets confused three. The authors have to make sure their simulation isn't introducing distortions that confuse their classifier.
Kai: They also used transfer learning from ImageNet one thousand to train these models. They say this initialization helps speed up convergence and makes the training more stable when they only have limited simulated events <ref:2609.25681#pg3>.
Mira: It leverages those pre-trained weights to get generic feature detectors like edges and curves, which then adapt quickly to the actual detector topologies. That’s a smart way to handle having limited training data for something so complex.
Lev: From an engineering standpoint, they tuned the selection efficiency by choosing an optimal cut value independently for each of the five Super-Kamiokande run periods, from SK-I through SK-V thirty-two. That shows they're being careful about consistency across all their data.
Kai: So to sum up the summary, the core idea is applying a sophisticated CNN with attention mechanisms and transfer learning to search for this specific trinucleon decay signal in Super-Kamiokande.
Mira: And while they didn't find a definitive signal, they established a lower limit on the decay lifetime at more than four point two times ten cubed years. That's an improvement over previous limits by more than six orders of magnitude.
Lev: So what this means for the world is that machine learning can be a powerful tool to enhance sensitivity in these large datasets when traditional methods might struggle with the sheer volume of data forty-seven.
Kai: It shows that using ML isn't just for classification; it’s being used to actually constrain physics beyond the Standard Model, connecting theory and experiment really tightly.
Mira: This result gives us a concrete benchmark for how sensitive these new ML techniques are when tested against theoretical predictions around one thousand thirty-three years three. It’s a solid piece of work for probing that physics.
Lev: For real hardware development, this study highlights that achieving high fidelity in event classification requires carefully managing systematic uncertainties like the pion final state interactions, which they quantified at about twenty-two to twenty-six percent thirty-two. That's a hard number for any detector design.
The paper's summary: Kai: So, we’re talking about how the researchers in this paper suggest ways to make their machine learning search even better than what they actually did.
Mira: They suggest enhancing event classification through a convolutional neural network architecture with an attention mechanism thirty-two. This means the model can focus on the most relevant parts of the Cherenkov ring patterns to improve how they classify events and suppress background.
Lev: That sounds like a big step for sensitivity, but we need to look at why that specific architecture works better for this physics problem than something simpler <ref:2609.25681#pg3>. What assumption are they making about the signal shape?
Kai: They’re using a modified MobileNetV3 structure with squeeze and excitation mechanisms for adaptive weighting thirty-four. That’s designed to reduce the number of parameters while still capturing complex spatial features, which is good for reducing computational overhead <ref:2609.25681#pg3>.
Mira: And they also use transfer learning from ImageNet one thousand initialization to accelerate convergence and improve stability when training with limited simulated events thirty-six. It gives the filters generic feature detectors like edges and curves right at the start <ref:2609.25681#pg3>.
Lev: But the next point is really about optimizing the cut value. They implement a scoring function that balances true positive and true negative rates, and they choose the optimal cut value where that score reaches its maximum thirty-two. It’s period-specific optimization <ref:2609.25681#pg3>.
Kai: That makes sense because the background environment changes across different operational periods of Super-Kamiokande, so tailoring the cut helps maintain consistent signal selection efficiency while keeping background rejection high thirty-two.
Mira: And they also compare their machine learning selections against full event selection using an event-by-event migration comparison thirty-two. This lets them determine if either selection contains the other, and in this case, they found that neither selection contains the other <ref:2609.25681#pg3>.
Lev: That’s a strong conclusion because it shows that these two selection methods are identifying different and complementary signal-background populations thirty-two. It validates the ML approach by showing it’s not just picking up noise; it's finding something distinct.
Kai: So the implication is that for future searches, we should expect to see this kind of optimization where the model learns from its performance across different operational phases.
Mira: And honestly, they are setting a very clear path forward by showing how ML can be used not just as a filter but as a tool to validate the selection process itself <ref:2609.25681#pg3>.
Lev: This gives us something concrete to work on for real hardware because we know we need that level of period-specific tuning to make these models reliable in practice, even if it adds complexity.
The paper's improvements: Kai: So we’re wrapping up our look at "Search for the sixteen O(ppp) to thirteen C pi+ pi+ e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques". We saw how they used machine learning to set a new lifetime limit against theoretical predictions.
Mira: Exactly. The core idea is that this trinucleon decay search, which looks for baryon number violation, can be done much more efficiently by using an AI model instead of just traditional counting methods.
Lev: For us on hardware, this means that if you build an experiment like Super-Kamiokande, you need to account for those pion final state interactions because they end up being about twenty-two to twenty-six percent of the uncertainty thirty-two. That’s a big modeling hurdle for any real detector design.
Kai: Right, so the whole point is that machine learning helps filter that massive amount of data from Super-Kamiokande in a way that traditional methods can’t handle alone, giving us this much tighter constraint.
Mira: It proves that connecting theory and experiment through these ML techniques really works when you have huge datasets to work with against rare signals. It gives us confidence in using AI to search for these elusive processes.
Lev: Yeah, but we still have that modeling uncertainty to tackle on the ground; the actual physics of those interactions is what will ultimately limit how precise we can get our experimental constraints on real hardware.
Kai: So, this paper shows a path forward where advanced pattern recognition can significantly enhance the sensitivity of nucleon decay searches. It’s a practical tool for high energy physics experiments right now.
Conclusion: Kai: So we’ve looked at "Search for the sixteen O(ppp) to thirteen C pi+ pi+ e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques" and found that they used a sophisticated convolutional neural network with attention mechanisms to search for nucleon decay <ref:2609.25681#pg3>.
Mira: And the key result is setting a lifetime limit of four point two times ten cubed years against a theoretical expectation around one thousand thirty-three years three <ref:2609.25681#pg1>. It's a solid constraint we can use to guide where we look next in the search for physics beyond the Standard Model.
Lev: For us on hardware, that limit sets a very high bar for what we need to measure in terms of sensitivity and background rejection on real detectors forty-seven <ref:2609.25681#pg1>. It tells us how much exposure we need to get a meaningful constraint.
Kai: The implication is that machine learning can be a powerful tool to enhance the sensitivity of nucleon decay searches by effectively filtering complex backgrounds in large datasets like Super-Kamiokande <ref:2609.25681#pg3>. It shows ML isn't just for classification; it’s for discovery itself.
Mira: This result proves that connecting theory and experiment through these machine learning techniques really works when you have huge datasets to work with against rare signals <ref:2609.25681#pg1>. It gives us confidence in using AI to search for those elusive processes.
Lev: Yeah, but we still have that modeling uncertainty to tackle on the ground; the actual physics of those interactions is what will ultimately limit how precise we can get our experimental constraints on real hardware <ref:2609.25681#pg3>.
Kai: So, this paper shows a path forward where advanced pattern recognition can significantly enhance the sensitivity of nucleon decay searches. It’s a practical tool for high energy physics experiments right now.
Mira: It’s definitely a solid piece of work that links those theoretical lifetime predictions directly to experimental limits through this ML approach <ref:2609.25681#pg1>.
Lev: We'll keep looking at how to reduce those modeling uncertainties on our side so that when we build the next generation of detectors, we can actually test these kinds of high-level analyses with confidence.
Kai: That’s all for this look at the "Search for the sixteen O(ppp) to thirteen C pi+ pi+ e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques" <ref:2609.25681#pg1,e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques>.
Mira: Thanks for listening. Next time we’ll be looking at how error correction is changing the game in quantum computing.
Kamioka Observatory Institute for Cosmic Ray Research University of Tokyo · Research Center for Cosmic Neutrinos Institute for Cosmic Ray Research University of Tokyo · Department of Theoretical Physics University Autonoma Madrid · Science Department Borough of Manhattan Community College City University of New York · Department of Physics Boston University · Department of Physics British Columbia Institute of Technology · School of Physical Sciences University Chinese Academy of Sciences · Institute for Universe and Elementary Particles Chonnam National University Institute for Space-Earth Environmental Research Nagoya University Department of Physics Gifu University GIST College Gwangju Institute of Science and Technology School of Physics and Astronomy University Glasgow Department of Physics California State University Dominguez Hills School of Mathematical and Physical Sciences University Sheffield August Chełkowski Institute of Physics Faculty of Science Tohoku University Department of Physics Oxford University Rutherford Appleton Laboratory Seoul National University Rutherford Appleton Laboratory STFC · University Tokyo Institutes for Advanced Study Kavli Institute for the Physics and Mathematics of the Universe
hep-ex
Submitted: 2026-09-22
Updated: 2026-11-12
Comments: 21 pages, 15 figures, 7 tables; accepted for publication in Phys. Rev. D
DOI: 10.1103/lydr-tkcd
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 98/100
The gist: Search for the 16O(ppp) →13 Cπ +π +e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques is a study that uses machine learning to search for nucleon decay…
Key concepts
- Nucleon Decay ($ riangle B = 3$)
- This refers to a hypothetical process where three nucleons (protons or neutrons) combine or transform into lighter particles, violating baryon number conservation by three units. This specific decay mode is considered the lowest-order baryon number violation allowed by certain theoretical symmetries and is a key target for physics beyond the Standard Model.
- Convolutional Neural Network (CNN) with Attention
- A deep learning architecture used to classify events in Super-Kamiokande data. The CNN analyzes spatial patterns in the detector response, while the attention mechanism helps the model focus on specific regions of these patterns that are most indicative of a rare decay signal, improving sensitivity over traditional methods.
- Final State Interactions (FSI)
- These are complex nuclear effects where particles produced during a decay interact with other nucleons inside the nucleus. The study simulates these interactions using models like NEUT cascade to accurately predict how the final particles will appear in the detector, which is crucial for correctly identifying a true signal.
- MobileNetV3 Architecture
- A specific type of CNN structure chosen for its efficiency and performance in event classification. It uses 'Squeeze-and-Excitation' mechanisms to adaptively weight features and a 'Bottleneck' structure to reduce data complexity, making the model effective even with limited simulated training data.
Terminology
Summary
Search for the 16O(ppp) →13 Cπ +π +e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques is a study that uses machine learning to search for nucleon decay, specifically focusing on the trinucleon decay mode in oxygen nuclei, which has implications for physics beyond the Standard Model The authors report a new partial lifetime limit of 4. p. 12 × 1032 years for the trinucleon decay mode 16O(ppp) →13 Cπ +π +e+ obtained from a search conducted using the Super-Kamiokande detector with 0. p. 1401 megaton-years of exposure across five operational periods This represents an improvement of six orders of magnitude over previous experimental constraints
Theoretical Motivation and Signal Definition
The study focuses specifically on processes with ∆B = 3, which can occur through dimension-15 operators and represent the lowest-order baryon number violation permitted by this symmetry The specific process investigated is trinucleon decay in oxygen nuclei, where 16O(ppp) undergoes decay into 13Cπ+π+e+ Theoretical predictions for such processes estimate their partial lifetime to be on the order of 1033 years This channel is considered an especially sensitive probe of baryon-parity-protected scenarios because it respects B−L conservation and can be embedded in unified constructions that address neutrino masses and the origin of the baryon asymmetry
Data Handling and Simulation Pipeline
The analysis utilized Monte Carlo (MC) simulations for both the trinucleon decay signal and atmospheric neutrino background for each SK detector period The simulation of the trinucleon decay takes into account Fermi motion, correlations between nucleons, and nuclear binding energy Pion-nucleon interactions within the nucleus (pion FSI) are simulated using the NEUT cascade model to address final state interactions Atmospheric neutrinos constitute the primary background, with dominant processes including charged-current single pion production (CC1π) and deep inelastic scattering (DIS)
Machine Learning Methodology
The analysis employs a convolutional neural network (CNN) incorporating an attention mechanism to enhance event classification and improve sensitivity The CNN architecture is a modified version of the MobileNetV3 architecture, which uses Squeeze-and-Excitation mechanisms for adaptive weighting and a Bottleneck structure for dimensional reduction The detector response data is projected into a 2D plane using a Mercator projection to create an input array of dimensions [3, 224, 224] for the CNN
Performance and Systematic Uncertainties
The trained CNN model was applied to pre-selected samples from all five SK periods, producing a distribution with two distinct peaks separating the signal from the background The optimal cut value is determined independently for each SK data-taking period (SK-I through SK-V) to ensure consistent signal selection efficiency while maintaining effective background rejection Systematic uncertainties are categorized into physics modeling and detector/reconstruction components, with the FSI of pions being the dominant systematic uncertainty affecting signal efficiency at approximately 22–26% The final 90% confidence-level upper limit on the decay rate is established by combining posterior distributions from all five SK run periods
Conclusion and Lifetime Limit
The search resulted in only one event remaining in the signal region, which was consistent with an expected background of approximately 1.33 events, showing no statistically significant excess Consequently, a lower limit on the trinucleon decay lifetime at τ > 4. p. 12 × 1032 years is established This result achieves more than a six-order-of-magnitude improvement over previous limits and reaches the theoretical model prediction for this ∆B = 3 mode
References
[1] A. D. Sakharov, “Violation of CP invariance, C asymmetry, and baryon asymmetry of the universe,” JETP Lett. p. 2 5, 24-27 (1967)
[2] H. Georgi and S. L. Glashow, “Unity of All Elementary Particle Forces,” Phys. p. 1 Rev. Lett. 32, 438 (1974)
[3] A. Takenaka et al., “Search for proton decay via p → e+π0 and p → µ+π0 with an enlarged fiducial volume in SuperKamiokande I-IV,” Phys. p. 2 Rev. D 102, 112011 (2020)
[4] K. S. Babu, I. Gogoladze, and K. Wang, “Gauged baryon parity and nucleon stability,” Phys. p. 2 Lett. B 570, 32-38 (2003)
[5] M. Agostini et al., “Search for tri-nucleon decays of 76Ge in GERDA,” Eur. p. 3 Phys. J. C 83, 778 (2023)
[6] S. I. Alvis et al., “Search for Tri-Nucleon Decay in the Majorana Demonstrator,” Phys. p. 2 Rev. D 99, 072004 (2019)
[7] Y. Fukuda et al. p. 0, “Super-Kamiokande and Other Detectors: A Case Study of Large Volume Cherenkov Neutrino Detectors,” in Instrumentation and Techniques in High Energy Physics (World Scientific, 2024), Chap. p. 2 8, pp. 251–290
[8] S. Mine, “The Super-Kamiokande and Other Detectors: A Case Study of Large Volume Cherenkov Neutrino Detectors,” in Instrumentation and Techniques in High Energy Physics (World Scientific, 2024), Chap. p. 2 8, pp. 251–290
[9] S. Yamada et al., “Commissioning of the New Electronics and Online System for the Super-Kamiokande Experiment,” IEEE Trans. p. 0 Nucl. Sci. 57, 428-432 (2010)
[10] H. Nishino, K. Awai, Y. Hayato, S. Nakayama, K. p. 0 Okumura, M. p. 0 Shiozawa, A. p. 0 Takeda, K. Ishikawa, A. Minegishi, and Y. Arai, Nucl. Instrum. Methods Phys Res. p. 2 p. 2, Sect a 610 710 (2009)
[11] K. Abe et al., “Calibration of the Super-Kamiokande detector,” Nucl. p. 1 Instrum. Methods Phys Res, Sect a 737, 253 (2014)
[12] D. Rein and L. M. Sehgal, “Neutrino-excitation of baryon resonances and single pion production,” Ann. p. 3 Phys. (N Y.) 133, 79 (1981)
[13] J. Gustafson et al., “Search for dinucleon decay into pions at SuperKamiokande,” Phys. p. 2 Rev. D 91, 072009 (2015)
[14] R. Matsumoto et al. p. 0, “Search for proton decay via p → µ+K0 in 0. p. 437 megaton-years exposure of Super-Kamiokande,” Phys. p. 1 Rev. D 106, 072003 (2022)
[15] K. Nakamura, S. p. 0 Hiramatsu, T. Kamae, H. Muramatsu, N. Izutsu, and Y. Watase, “The reaction 12C(e, e'p) at 700 MeV and DWIA analysis,” Nucl. Phys. A 268, 381 (1976)
[16] M. G. Mayer and J. H. D. Jensen, Elementary Theory of Nuclear Shell Structure (Wiley, New York, 1955)
[17] T. Yamazaki and Y. Akaishi, “Nuclear medium effects on invariant mass spectra of hadrons decaying in nuclei,” Phys. p. 5 Lett. B 453, 1 (1999)
[18] R. D. Woods and D. p. 4 S. Saxon, “Diffuse Surface Optical Model for Nucleon-Nuclei Scattering,” Phys. p. 4 Rev. 95, 577 (1954)
[19] M. Shiozawa (Super-Kamiokande Collaboration), Nucl. p. 0 Instrum. Methods Phys Res., Sect a 433, 240 (1999)
[20] Y. Hayato, “NEUT,” Nucl. p. 0 Phys. B, Proc. Suppl 112, 171 (2002)
[30] E. R. Davies, Machine Vision: Theory, Algorithms, Practicalities, 2nd ed. p. 5, Academic Press (San Diego, 1997)
[31] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner, Gradient-based learning applied to document recognition, Proceedings of the IEEE, 86(11):2278-2324 (1998)
[32] Andrew Howard, Mark Sandler, Grace Chu, LiangChieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le., Hartwig Adam. Searching for MobileNetV3. p. 8 arXiv:1905. p. 002244 [cs.CV], 2019
[33] J. Hu, L. Shen, and G. Sun, “Squeeze-and-Excitation Networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 7132-7141 (2018)
[34] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: Inverted Residuals and Linear Bottlenecks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 4510-4520 (2018)
[35] A. Paszke et al., “PyTorch: An Imperative Style, HighPerformance Deep Learning Library,” Advances in Neural Information Processing Systems 32, 8024-8035 (2019)
[36] R. Wightman, “PyTorch Image Models (timm): MobileNetV3 pretrained on ImageNet-1k,” https://huggingface. p. 8co/timm/mobilenetv3 large 100. p. 8railin1k
[37] K. Abe et al., “Improved constraints on neutrino mixing from the T2K experiment with 3. p. 113 × 1021 protons on target,” Phys. p. 2 Rev. D 103, 112008 (2021)
[38] K. Abe et al., “Atmospheric neutrino oscillation analysis with external constraints in Super-Kamiokande I-IV,” Phys. p. 5 Rev. D 97, 072001 (2018)
[39] T. Wester et al. p. 0, “Atmospheric neutrino oscillation analysis with neutron tagging and an expanded fiducial volume in SuperKamiokande I–V,” Phys. p. 5 Rev. D 109, 072014 (2024)
[41] Department of Physics, Oxford University, Oxford, OX1 3PU, United Kingdom
[43] Department of Physics and Astronomy, Seoul National University, Seoul 151-742 (Korea)
[44] School of Mathematical and Physical Sciences, University of Sheffield, S3 7RH (United Kingdom)
[47] STFC, Rutherford Appleton Laboratory, Harwell Oxford, and Daresbury Laboratory, Warrington OX11 0QX (UK)
[48] Department of Physics, Sungkyunkwan University, Suwon 440-746 (Korea)
[50] Department of Physics, University of Tokyo, Bunkyo, Tokyo 113-0033 (Japan)
[51] Kavli Institute for the Physics and Mathematics of the Universe (WPI), The University of Tokyo Institutes for Advanced Study, University of Tokyo, Kashiwa, Chiba 277-8583 (Japan)
[53] Department of Physics and Astronomy, Faculty of Science and Technology, Tokyo University of Science, Noda, Chiba 278-8510 (Japan)
[54] Faculty of Science, University of Toyama, Toyama City (Japan)
[55] TRIUMF, 4004 Wesbrook Mall, Vancouver BC V6T2A3 (Canada)
[57] Faculty of Physics, University of Warsaw, Warsaw 02-093 (Poland)
[58] Department of Physics, University of Warwick, Coventry CV4 7AL (UK)
[59] Department of Physics, University of Winnipeg, MB R3J 3L8 (Canada)
[60] Department of Physics, Yokohama National University, Yokohama Kanagawa 240-8501 (Japan)
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
-
Enhanced event classification via CNN architecture with an attention mechanism enables
the model to focus on the most relevant regions of Cherenkov ring patterns—to enhance event classification.
This allows for improved identification of event signatures affected by FSI and better background suppression. -
Development of a cylindrical-to-planar projection mapping procedure preserves topological properties essential for pattern recognition, as
this transformation preserves the topological properties of Cherenkov ring patterns, which is essential for subsequent pattern recognition.
This geometric consistency ensures that distortions are learned consistently by the network rather than introducing data/MC bias. -
Training CNN models using transfer learning from ImageNet1k accelerates convergence and improves stability given limited simulated events, as
This transfer-learning initialization provides the convolutional filters with generic low-level feature detectors (edges, curves, and intensity gradients) that are also relevant for Cherenkov-ring patterns.
This leverages pre-trained weights to rapidly adapt to detector-specific topologies. -
Implementation of a scoring function balancing true positive rate and true negative rate allows for period-specific optimization of the CNN cut, as
The optimal cut, indicated by the green dashed vertical line, is chosen at the machine learning output value where this Score function reaches its maximum.
This ensures consistent signal selection efficiency while maintaining effective background rejection across all operational phases. -
Utilization of event-by-event migration comparison between CNN and full APFit selections provides a method to determine if
either selection contains the other,
asThe sizable exclusive populations in both directions show that neither selection contains the other.
This yields a conclusion thatthe two selections identify different and complementary signal-MC populations.
Sources
Related papers
- DrSai: an agentic AI system toward autonomous physics analysis workflows at BESIII
- CANTON- mu Proposal: A Next-Generation Muon g-2 Measurement at Sub-0.1 ppm Precision
- Search for dark photon dark matter near 1,meV using a superconducting mixer
- ANUBIS detector and its sensitivity to neutral long-lived particles
- How to scale your HEP ML models: A recipe for robust architecture comparisons at scale
- DMRadio-Core: A new approach for GUT-scale axion searches