Machine learning assisted nu/nu bar discrimination in atmospheric neutrino events for CPV sensitivity in a 20-kton liquid scintillator detector

arXiv:2611.11474 · hep-ex, hep-ph · Submitted 2026-11-08 · 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: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Machine learning assisted nu/nu bar discrimination in atmospheric neutrino events for CPV sensitivity in a 20-kton liquid scintillator detector".

Mira: Here's the gist of the paper. Machine learning based statistical neutrino-antineutrino separation can recover part of the CP violation sensitivity lost in charge blind detectors Atmospheric neutrinos span a broad range of energies and baselines, offering a complementary probe of leptonic CP violation The XGBoost classifier achieves an AUC of approximately 0.83 for both neutrino flavors Under the square root approximation, peak CP violation sensitivity improves from 1.49 σ to 1.96 σ for normal mass ordering and from 1.14 σ to 1.61 σ for inverted mass ordering Machine learning based statistical neutrino-antineutrino separation can recover part of the CP violation sensitivity lost in charge blind detectors Simulation and Feature Engineering The event generation involves creating neutrino events with GENIE using the Honda flux tables and then reweighting them using OscProb three flavor oscillation probabilities incorporating Earth matter effects and NuFIT parameters Detector response is approximated by smearing lepton and hadronic energies, neutrino direction, and neutron multiplicity The reconstructed inelasticity is computed as y = Ehad/(Ehad +El) The neutron multiplicity, nf n, is modeled by applying a 90% binomial tagging efficiency to the true final state neutron count Model Selection and Performance Three machine learning classifiers were trained on the reconstructed features y, nf n, and El using independent datasets of approximately 4.0×10 6 (νµ) and 2.0×10 6 (νe) CC events Logistic regression provides a linear baseline, while XGBoost and the MLP can capture nonlinearities XGBoost was selected for the analysis based on its classification performance and computational efficiency The reported percentages represent the relative feature importance within the trained model, not a direct measure of the independent physical information carried by each observable Sensitivity Enhancement The ML assisted method improves the peak CPV sensitivity ∆χ 2 CPV from 2.22 (1.49σ) to 3.86 (1.96σ) for NO and from 1.29 (1.14σ) to 2.58 (1.61σ) for IO as compared to the charge blind case The ideal separation scenario provides an upper benchmark with perfect ν/ν¯ identification producing peak ∆χ 2 CPV values of 6.98 (NO) and 5.43 (IO) The ML assisted analysis enhances sensitivity relative to the charge blind baseline at all CP violating values of δCP, partially recovering the lost CP sensitivity, though the magnitude depends on δCP and mass ordering These results are preliminary because the detector response is approximated by parameterized smearing and a simplified neutron tagging model More realistic modeling of neutron detection, event reconstruction, backgrounds, and correlated systematics is required to assess the robustness of this approach The study shows that ML based statistical discrimination is a promising way to improve CPV sensitivity of a large LS detector REFERENCES

1: T.

Kai: Birkenfeld and A.

Mira: Stahl.

Kai: J.

Paper summary: Kai: The paper, "Machine learning assisted nu/nu bar discrimination in atmospheric neutrino events for CPV sensitivity in a twenty-kton liquid scintillator detector", is testing the idea that machine learning can statistically separate neutrino and antineutrino events <ref:2611.11474#pg1,discrimination in atmospheric neutrino events for CPV sensitivity in a 20-kton>. They use datasets of about four point zero times ten to the six muon neutrinos and two point zero times ten to the six electron neutrinos charged current events.

Mira: The main claim they make is that an XGBoost classifier achieves an area under the receiver operating characteristic curve, or AUC, of approximately zero point eight three for both neutrino flavors. They also look at a linear baseline with logistic regression, and then they compare that against the nonlinear capabilities of XGBoost and a multilayer perceptron.

Lev: The feature importance table shows that neutron multiplicity is the dominant model derived feature, contributing approximately seventy seven percent for muon neutrinos and seventy nine percent for electron neutrinos. That tells us what observable has the most weight in this separation process based on their training.

Kai: They look at how this separation improves the peak CP violation sensitivity. Under the square root approximation, they see it improving from one point four nine sigma to one point nine six sigma for normal mass ordering, and from one point one four sigma to one point six one sigma for inverted mass ordering.

Mira: This is compared against the charge blind case where they get sensitivity of only two point two two times two decimal two, which is about one point four nine sigma for normal ordering and one point zero nine times one point one four sigma for inverted ordering. The comparison shows a significant improvement in the expected signal strength.

Lev: They also look at the ideal separation scenario, where you get perfect neutrino-antineutrino discrimination, and they find peak sensitivity values of six point nine eight times two decimal two for normal ordering and five point four three times three decimal one for inverted ordering. That’s the upper bound they are aiming for with perfect separation.

Kai: The paper is showing that this machine learning assisted method can recover some of the CP violation sensitivity that would otherwise be lost in detectors without magnetic fields. The data they use is based on two point zero times ten to the six muon neutrinos and two point zero times ten to the six electron neutrinos CC events partitioned into training, validation, calibration, and testing sets.

Mira: It’s a promising result because it shows that these statistical techniques can play a role in extracting CP violation information even when the detector setup itself has limitations regarding direct charge identification. This result comes from the study of atmospheric neutrinos spanning a broad range of energies and baselines.

Lev: The method they use involves training three different classifiers, logistic regression, XGBoost, and an MLP on the features y, n n, and E l. XGBoost was chosen for its classification performance and computational efficiency when compared to the other two models.

Conclusion: Kai: So, looking at the paper "Machine learning assisted nu/nu bar discrimination in atmospheric neutrino events for CPV sensitivity in a twenty-kton liquid scintillator detector", what we see is that machine learning based statistical separation helps recover some of the CP violation sensitivity lost in charge blind detectors <ref:2611.11474#pg1,discrimination in atmospheric neutrino events for CPV sensitivity in a 20-kton>. The authors are Dek Hao Liew, Mohammad Nizam, Shu Tian Eu, and Siew Yan Hoh.

Mira: The key thing here is that they've shown this improvement depends on the CP violating values of delta CP and the mass ordering of neutrinos. They recovered a peak sensitivity of one point nine six sigma for normal mass ordering and one point six one sigma for inverted mass ordering when using their XGBoost classifier.

Lev: It’s important to remember that these results are preliminary because the detector response is approximated by parameterized smearing, and they need more realistic modeling of neutron detection, event reconstruction, backgrounds, and correlated systematics to truly assess the robustness of this approach.

Kai: Exactly. They aren't just looking at a perfect scenario. This study shows that machine learning based statistical discrimination is a promising way to improve CPV sensitivity for a large detector like the LS detector they are studying by using atmospheric neutrinos as their probe.

Mira: It suggests that even with limitations in the hardware, sophisticated statistical methods on reconstructed features can still be very useful for probing fundamental physics like CP violation.

Lev: We need to see how this translates when you put it on a real machine where you have to account for all those systematic uncertainties they mentioned. That’s the next step before we can say much more about making this method viable for real hardware.

Dek Hao Liew, Mohammad Nizam, Shu Tian Eu, Siew Yan Hoh

Department of Physics, Xiamen University Malaysia · Santa Cruz Institute for Particle Physics, University of California Santa Cruz

hep-ex, hep-ph

Submitted: 2026-11-08

Updated: 2026-11-08

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

Importance score: 77/100

The gist: Machine learning based statistical neutrino-antineutrino separation can recover part of the CP violation sensitivity lost in charge blind detectors Atmospheric neutrinos span a broad range of…

Key concepts

Atmospheric Neutrinos
These are neutrinos produced when cosmic rays interact with the Earth's atmosphere. They come in a wide range of energies and travel different paths, making them a valuable tool to study leptonic CP violation.
XGBoost Classifier
This is a machine learning model used to distinguish between neutrino and antineutrino events based on reconstructed features like inelasticity and neutron multiplicity. It was chosen for its good classification performance and computational efficiency.
CP Violation Sensitivity ($ riangle ext{CPV}$)
This measures how well a detector can distinguish between neutrino oscillations under different CP violation scenarios. The ML method improved this sensitivity, meaning the detector could better probe the fundamental properties of neutrinos.
Inelasticity ($y$)
Inelasticity is a calculated value derived from the reconstructed energies of hadronic and lepton components of an event. It helps characterize the interaction type, which is then used as an input feature for the machine learning model.

Terminology

Summary

Machine learning based statistical neutrino-antineutrino separation can recover part of the CP violation sensitivity lost in charge blind detectors

Atmospheric neutrinos span a broad range of energies and baselines, offering a complementary probe of leptonic CP violation The XGBoost classifier achieves an AUC of approximately 0.83 for both neutrino flavors Under the square root approximation, peak CP violation sensitivity improves from 1. p. 049 σ to 1.96 σ for normal mass ordering and from 1. p. 0 p. 014 σ to 1.61 σ for inverted mass ordering Machine learning based statistical neutrino-antineutrino separation can recover part of the CP violation sensitivity lost in charge blind detectors

Simulation and Feature Engineering

The event generation involves creating neutrino events with GENIE using the Honda flux tables and then reweighting them using OscProb three flavor oscillation probabilities incorporating Earth matter effects and NuFIT parameters Detector response is approximated by smearing lepton and hadronic energies, neutrino direction, and neutron multiplicity The reconstructed inelasticity is computed as y = Ehad/(Ehad +El) The neutron multiplicity, nf n, is modeled by applying a 90% binomial tagging efficiency to the true final state neutron count

Model Selection and Performance

Three machine learning classifiers were trained on the reconstructed features y, nf n, and El using independent datasets of approximately 4.0×10 6 (νµ) and 2. p. 00×10 6 (νe) CC events Logistic regression provides a linear baseline, while XGBoost and the MLP can capture nonlinearities XGBoost was selected for the analysis based on its classification performance and computational efficiency The reported percentages represent the relative feature importance within the trained model, not a direct measure of the independent physical information carried by each observable

Sensitivity Enhancement

The ML assisted method improves the peak CPV sensitivity ∆χ 2 CPV from 2.22 (1.49σ) to 3. p. 286 (1.96σ) for NO and from 1. p. 229 (1.14σ) to 2. p. 258 (1.61σ) for IO as compared to the charge blind case The ideal separation scenario provides an upper benchmark with perfect ν/ν¯ identification producing peak ∆χ 2 CPV values of 6. p. 298 (NO) and 5.43 (IO)

Conclusion and Limitations

The ML assisted analysis enhances sensitivity relative to the charge blind baseline at all CP violating values of δCP, partially recovering the lost CP sensitivity, though the magnitude depends on δCP and mass ordering These results are preliminary because the detector response is approximated by parameterized smearing and a simplified neutron tagging model More realistic modeling of neutron detection, event reconstruction, backgrounds, and correlated systematics is required to assess the robustness of this approach The study shows that ML based statistical discrimination is a promising way to improve CPV sensitivity of a large LS detector

REFERENCES

[1] T. Birkenfeld and A. p. 3 Stahl. J. Instrum. 21 (2026), P04015

[2] X. He et al. Chin. Phys. C 50 (2026), 093001

[3] J. Liu et al. Phys Rev D 112 (2025), 012018

[4] M Honda et al. Phys Rev D 92 (2015), 023004

[5] Z Yang et al. Phys Rev D 109 (2024), 052005

[6] W Guo et al. J Phys Conf Ser. 888 (2017), 012205

[7] S Choubey and T Ohlsson. Phys Lett B 739 (2014), 357–364

[8] R Wendell et al. Phys Rev D 81 (2010), 092004

[9] M C Gonzalez-Garcia and M Maltoni. Phys Rev D 70 (2004), 033010

--- Page 1 ---

Machine learning assisted ν/ν¯ discrimination in atmospheric neutrino events for CPV sensitivity in a 20-kton liquid scintillator detector Dek Hao Liew1, Mohammad Nizam1,2,∗, Shu Tian Eu1, Siew Yan Hoh1 Department of Physics, Xiamen University Malaysia, Sepang, Selangor, Malaysia 2Santa Cruz Institute for Particle Physics Santa Cruz Institute for Particle Physics University of California Santa Cruz Santa Cruz California USA ∗Corresponding author: mohammad.nizam@xmu.edu.my Abstract Purpose: Atmospheric neutrinos span a broad range of energies and baselines, offering a complementary probe of leptonic CP violation However, a JUNO-like liquid scintillator detector without a magnetic field cannot directly separate neutrino and antineutrino interactions, reducing CP violation sensitivity This study investigates whether machine learning based statistical separation can improve CP violation sensitivity in a 20-kton detector Methods: νµ and νe charged current events are generated with GENIE (using Honda flux) in a JUNO-like detector geometry and are reweighted using OscProb three flavor oscillation probabilities incorporating Earth matter effects and NuFIT parameters Detector response is approximated by smearing lepton and hadronic energies, neutrino direction, neutron multiplicity. p. 0 Logistic regression, XGBoost, and multilayer perceptron (MLP) classifiers are trained on inelasticity, neutron multiplicity, and charged lepton energy with XGBoost selected for the analysis CP violation sensitivity is evaluated for a 200 kton-year exposure using a Poisson likelihood with systematic nuisance parameters and marginalization over five oscillation parameters within their respective 3σ ranges Results: XGBoost achieves an area under the receiver operating characteristic curve (AUC) of approximately 0. p. 083 for both flavor samples Under the square root approximation, peak CP violation sensitivity improves from 1. p. 049 σ to 1.96 σ for normal mass ordering and from 1.14 σ to 1.61 σ for inverted mass ordering Conclusion: Machine learning based statistical neutrino-antineutrino separation can recover part of the CP violation sensitivity lost in charge blind detectors These preliminary results motivate further studies incorporating validated detector response and systematic uncertainties Keywords: Atmospheric neutrinos; liquid scintillator; machine learning; CP violation (CPV) Acknowledgments: This work was supported by Xiamen University Malaysia (XMUM) under the XMUM Research Fund (XMUMRF) grant number: XMUMRF/2025-C15/IPHY/0007. p. 0

--- Page 2 ---

True neutrino directions are Gaussian smeared to obtain an energy dependent angular resolution of 5◦–28◦ [5] From the smeared energies, the reconstructed inelasticity is computed as y = Ehad/(Ehad +El) Events with 0.1 ≤ Erec ≤ 10 GeV are retained for the analysis 1 True neutrino directions are Gaussian smeared to obtain an energy dependent angular resolution of 5◦–28◦ [5] From the smeared energies, the reconstructed inelasticity is computed as y = Ehad/(Ehad +El) Events with 0. p. 11 ≤ Erec ≤ 10 GeV are retained for the analysis

--- Page 3 ---

Table 1 shows that nf n provides the largest contribution to the XGBoost classification, followed by y and El. The reported percentages represent the relative feature importance within the trained model, not a direct measure of the independent physical information carried by each observable

--- Page 4 ---

The classifier achieves an AUC of approximately 0.83 for both neutrino flavors Neutron multiplicity is the dominant model derived feature (contributing approximately 77% and 79% for the muon and electron flavor samples, respectively)

--- Page 5 ---

Ordering Charge blind ML assisted Ideal separation NO 2.22 (1.49) 3.86 (1.96) 6.98 (2.64) IO 1.29 (1.14) 2.58 (1.61) 5 cubed

--- Page 5 ---

Table 2: Peak CPV sensitivity for an exposure of 200 kton-year Values in parentheses are the approximate significance

--- Page 5 ---

Ordering Charge blind ML assisted Ideal separation NO Charge blind ML assisted Ideal separation NO IO Charge blind ML assisted Ideal separation IO

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. Machine learning assisted statistical neutrino-antineutrino separation can recover part of the CP violation sensitivity lost in charge blind detectors by training classifiers like XGBoost on reconstructed features such as inelasticity, neutron multiplicity, and charged lepton energy. This improved system can statistically sort events into enriched neutrino-like and antineutrino-like categories to enhance CPV sensitivity.

  2. The AI system can achieve a peak approximate CPV sensitivity improvement from 1.49 sigma to 1.96 sigma for normal mass ordering in a 20kton detector, which is more sensitive than the charge blind case of 1.49 sigma, as stated in Table 2. This allows the AI to provide a preliminary benchmark for ML assisted atmospheric neutrino CPV searches under specific oscillation scenarios.

  3. The system can utilize feature importance analysis to identify that Neutron multiplicity is the dominant model derived feature (contributing approximately 77% and 79% for the muon and electron flavor samples, respectively). This allows researchers to prioritize which detector response variables are most valuable for future ML training on realistic datasets.

  4. The AI system can be configured to perform a Poisson likelihood ratio statistic calculation using nuisance parameters [1] to evaluate CPV sensitivity, incorporating seven systematic uncertainties such as flux normalization (20%), cross section normalization (10%), and zenith angle tilt (5%). This enables the system to provide a more robust CPV sensitivity estimate by marginalizing over oscillation parameters within their 3 sigma ranges.

Related papers