pith. sign in

arxiv: 2606.03745 · v1 · pith:ZV2OA6MLnew · submitted 2026-06-02 · ✦ hep-ph · cs.LG· hep-ex· physics.data-an

Predicting the Neutrino Mass Ordering Using Neural Networks

T.J.C. Bezerra , L. Asquith , E. Bannister , W. Shorrock This is my paper

Pith reviewed 2026-06-28 09:17 UTC · model grok-4.3

classification ✦ hep-ph cs.LGhep-exphysics.data-an
keywords neutrino mass orderingneural networkslong-baseline experimentsneutrino oscillationsmachine learningmass hierarchy
0
0 comments X

The pith

A neural network classifier matches the performance of standard statistical fits when determining neutrino mass ordering from synthetic long-baseline data.

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

The paper tests whether a feed-forward neural network can classify neutrino mass ordering as normal or inverted using simulated data from long-baseline oscillation experiments. Synthetic datasets incorporate three-flavour probabilities, matter effects, and statistical fluctuations to train and evaluate the network. Performance is measured against chi-squared and likelihood methods via metrics such as ROC curves, with results showing comparable discrimination power. This suggests the network can serve as an independent cross-check for an open question that current data cannot yet resolve. The approach is positioned as extensible to include systematics or joint parameter inference.

Core claim

The neural network achieves performance comparable to conventional fits for the scenarios studied, providing a flexible, independent cross-check of established analyses for neutrino mass ordering determination.

What carries the argument

Feed-forward neural-network classifier trained on synthetic long-baseline datasets that include three-flavour oscillation probabilities, matter effects, and statistical fluctuations; it outputs a classification score for normal versus inverted ordering.

If this is right

  • Operating points on the classifier can be chosen to favour either higher purity or higher efficiency in mass-ordering assignment.
  • The same framework can incorporate systematic uncertainties in future extensions.
  • Joint inference of multiple oscillation parameters becomes feasible within the neural-network approach.
  • The method offers a potential pedagogical example for applying machine learning to neutrino oscillation problems.

Where Pith is reading between the lines

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

  • If the network learns correlations missed by traditional fits, it could improve sensitivity in experiments where parameter degeneracies are severe.
  • Retraining on more detailed detector simulations might reveal whether the comparable performance persists under realistic backgrounds and efficiencies.
  • The classifier could be combined with existing likelihood analyses to produce ensemble predictions that reduce reliance on any single method.

Load-bearing premise

The synthetic datasets used for training and testing are representative enough of real experimental conditions to make the performance comparison meaningful.

What would settle it

Applying the trained network to actual data from a long-baseline experiment and checking whether its mass-ordering discrimination matches or exceeds the chi-squared result on the same dataset.

Figures

Figures reproduced from arXiv: 2606.03745 by E. Bannister, L. Asquith, T.J.C. Bezerra, W. Shorrock.

Figure 1
Figure 1. Figure 1: The probability of muon-to-electron neutrino oscillations for neutrinos versus antineutrinos with the NOvA [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Oscillated energy spectra of νµ, ν¯µ, νe, and ν¯e generated using the parameter values in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Difference in chi-squared, ∆χ 2 , as a function of ∆m2 32 for both mass orderings at the ambiguity point (δCP = π/2 for NO and δCP = 3π/2 for IO). For each test hypothesis, ∆χ 2 is computed as the difference between the chi-squared value of the test histogram and that of a reference histogram generated under NO with ∆m2 32 = 2.45×10−3 eV2 . The minimum value under the IO assumption is ∆χ 2 min = 0.166, whi… view at source ↗
Figure 4
Figure 4. Figure 4: Left: The oscillated energy spectrum (orange) is obtained using the simplified two-flavour survival probabil [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results of NN-based MO classification using the training sample. Top: output probability of the NN [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between the NN model and standard LLH minimisation using an independent test sample of [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: ROC curve of the NN model applied to a test sample and standard LLH minimisation point. The data used to [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
read the original abstract

Determining the neutrino mass ordering remains a central open problem in particle physics. While next-generation long-baseline experiments are expected to resolve this question, current data provide limited sensitivity because the spectral differences between normal and inverted ordering are subtle and entangled with parameter degeneracies. We investigate a machine-learning strategy for mass-ordering determination using a feed-forward neural-network classifier trained on synthetic long-baseline datasets generated with three-flavour oscillation probabilities, matter effects, and statistical fluctuations. We evaluate the classifier against standard $\chi^2$ and $\log\mathcal{L}$ approaches using common discrimination metrics, including receiver-operating-characteristic curves, to quantify sensitivity and to illustrate how operating points can be selected to prioritise purity or efficiency. We find that the neural network achieves performance comparable to conventional fits for the scenarios studied, providing a flexible, independent cross-check of established analyses. The framework can be extended to incorporate systematic uncertainties and to explore joint inference of oscillation parameters, and it may also serve as a pedagogical tool for introducing machine-learning methods in neutrino physics.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The manuscript investigates a feed-forward neural-network classifier for determining neutrino mass ordering (normal vs. inverted) trained on synthetic long-baseline oscillation datasets generated from three-flavour probabilities including matter effects and Poisson statistical fluctuations. Performance is compared to standard χ² and log-likelihood fits via ROC curves and related discrimination metrics, with the conclusion that the NN achieves comparable results for the scenarios studied and can serve as a flexible, independent cross-check.

Significance. If the central comparison holds under the stated conditions, the work supplies a controlled proof-of-concept for machine-learning methods as a complementary tool in neutrino oscillation phenomenology, with noted extensibility to systematics and potential pedagogical utility. The idealized synthetic-data setting allows clean isolation of statistical effects but restricts immediate relevance to real experiments.

major comments (2)
  1. [Abstract] Abstract: the assertion that the neural network 'achieves performance comparable to conventional fits' is presented without any quantitative metrics (AUC values, specific ROC operating points, error estimates, or training hyperparameters), rendering the central claim unverifiable from the stated text.
  2. [Data-generation description (likely §2–3)] Data-generation description (likely §2–3): the synthetic events incorporate only oscillation probabilities, matter effects, and statistical fluctuations; detector smearing, energy-scale uncertainties, flux normalizations, and other systematics that must be marginalized in actual χ² analyses are omitted. This choice means the reported ROC performance does not test robustness against the dominant uncertainties of established methods, weakening the claim of an 'independent cross-check of established analyses'.
minor comments (1)
  1. [Abstract] Abstract and results section: include at least one table or figure caption with explicit numerical performance values (e.g., AUC or efficiency at fixed purity) to support the comparability statement.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. We respond to each major comment below and indicate planned revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that the neural network 'achieves performance comparable to conventional fits' is presented without any quantitative metrics (AUC values, specific ROC operating points, error estimates, or training hyperparameters), rendering the central claim unverifiable from the stated text.

    Authors: We agree that the abstract would benefit from quantitative metrics to support the comparability statement. In the revised version we will insert specific AUC values (extracted from the ROC curves already shown in the body), a note on selected operating points, and the main training hyperparameters. This change will make the central claim directly verifiable from the abstract text. revision: yes

  2. Referee: [Data-generation description (likely §2–3)] Data-generation description (likely §2–3): the synthetic events incorporate only oscillation probabilities, matter effects, and statistical fluctuations; detector smearing, energy-scale uncertainties, flux normalizations, and other systematics that must be marginalized in actual χ² analyses are omitted. This choice means the reported ROC performance does not test robustness against the dominant uncertainties of established methods, weakening the claim of an 'independent cross-check of established analyses'.

    Authors: The manuscript is explicitly framed as a controlled proof-of-concept that isolates statistical discrimination power; the absence of detector smearing and systematic uncertainties is stated in §§2–3 and is intentional for this scope. We acknowledge that the present results therefore cannot demonstrate robustness against the dominant experimental uncertainties. In revision we will expand the discussion section to restate this limitation clearly, moderate the phrasing of the 'independent cross-check' claim to reflect the idealized statistical setting, and reiterate the extensibility to systematics already noted in the abstract. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical comparison on synthetic data is self-contained

full rationale

The paper generates synthetic long-baseline datasets from three-flavour oscillation probabilities (including matter effects and Poisson fluctuations), trains a feed-forward NN classifier to discriminate mass ordering, and reports ROC-based performance metrics that are compared directly to standard χ² and logℒ fits on the same held-out synthetic events. No load-bearing step reduces a claimed prediction to a quantity defined by the same fit, no self-citation chain is invoked to justify uniqueness or an ansatz, and the central claim remains an empirical observation on independently generated test data rather than a definitional identity. The enumerated circularity patterns are absent; the result is therefore scored at the low end of the allowed range.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated beyond the standard three-flavour oscillation framework used to generate the training data.

axioms (1)
  • domain assumption Three-flavour neutrino oscillation probabilities with matter effects accurately describe the synthetic data generation process
    Invoked to produce the training and test datasets

pith-pipeline@v0.9.1-grok · 5725 in / 1102 out tokens · 35526 ms · 2026-06-28T09:17:58.660556+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

25 extracted references · 1 linked inside Pith

  1. [1]

    Suekane.Neutrino oscillations: a practical guide to basics and applications

    F. Suekane.Neutrino oscillations: a practical guide to basics and applications. Springer Tokyo, 2015

    2015

  2. [2]

    Qian and P

    X. Qian and P. V ogel. Neutrino mass hierarchy.Progress in Particle and Nuclear Physics, 83:1–30, 2015

    2015

  3. [3]

    Agostini et al

    M. Agostini et al. Comprehensive measurement of pp-chain solar neutrinos.Nature, 562:505–510, 2018

    2018

  4. [4]

    J. D. Vergados, H. Ejiri, and F. Šimkovic. Neutrinoless double beta decay and neutrino mass.International Journal of Modern Physics E, 25:1630007, 2016

    2016

  5. [5]

    Kajino, G.J

    T. Kajino, G.J. Mathews, and T. Hayakawa. Neutrinos in core-collapse supernovae and nucleosynthesis.Journal of Physics G: Nuclear and Particle Physics, 41:044007, 2014

    2014

  6. [6]

    Lesgourgues and S

    J. Lesgourgues and S. Pastor. Neutrino mass from cosmology.Advances in High Energy Physics, 2012:608515, 2012

    2012

  7. [7]

    Wolfenstein

    L. Wolfenstein. Neutrino oscillations in matter.Phys. Rev. D, 17:2369, 1978

    1978

  8. [8]

    Mikheyev and A.Y

    S.P. Mikheyev and A.Y . Smirnov. Resonance amplification of oscillations in matter and spectroscopy of solar neutrinos.Sov. J. Nucl. Phys., 42:913–917, 1985. 10 APREPRINT- JUNE3, 2026

    1985

  9. [9]

    Petcov and M

    S.T. Petcov and M. Piai. The lma msw solution of the solar neutrino problem, inverted neutrino mass hierarchy and reactor neutrino experiments.Physics Letters B, 533:94–106, 2002

    2002

  10. [10]

    Ahn et al

    M.H. Ahn et al. Measurement of neutrino oscillation by the K2K experiment.Phys. Rev. D, 74:072003, 2006

    2006

  11. [11]

    Adamson et al

    P. Adamson et al. Electron neutrino and antineutrino appearance in the full MINOS data sample.Phys. Rev. Lett., 110:171801, 2013

    2013

  12. [12]

    Joint neutrino oscillation analysis from the T2K and NOvA experiments.Nature, 646:818–824, 2025

    NOvA and T2K Collaborations. Joint neutrino oscillation analysis from the T2K and NOvA experiments.Nature, 646:818–824, 2025

    2025

  13. [13]

    Long-Baseline Neutrino Facility (LBNF) and Deep Underground Neutrino Experiment (DUNE) Conceptual Design Report V olume 2: The Physics Program for DUNE at LBNF

    R.Acciarri et al. Long-Baseline Neutrino Facility (LBNF) and Deep Underground Neutrino Experiment (DUNE) Conceptual Design Report V olume 2: The Physics Program for DUNE at LBNF. 2015

    2015

  14. [14]

    Abe et al

    K. Abe et al. Hyper-Kamiokande Design Report. 2018

    2018

  15. [15]

    Ayres et al

    D.S. Ayres et al. The NOvA technical design report. Technical Report FERMILAB-DESIGN-2007-01, Fermilab National Accelerator Laboratory, 2007

    2007

  16. [16]

    R. L. Talaga et al. PVC Extrusion Development and Production for the NOvA Neutrino Experiment.Nucl. Instrum. Meth. A, 861:77–89, 2017

    2017

  17. [17]

    Navas et al

    S. Navas et al. Review of particle physics.Phys. Rev. D, 110(3):030001, 2024. With 2025 update

    2024

  18. [18]

    Abubakar et al

    S. Abubakar et al. Precision measurement of neutrino oscillation parameters with 10 years of data from the NOvA experiment.Phys. Rev. Lett., 136:011802, 2026

    2026

  19. [19]

    Acero et al

    M.A. Acero et al. Improved measurement of neutrino oscillation parameters by the NOvA experiment.Phys. Rev. D, 106:032004, 2022

    2022

  20. [20]

    Denton and S.J

    P.B. Denton and S.J. Parke. Fast and accurate algorithm for calculating long-baseline neutrino oscillation proba- bilities with matter effects.Phys. Rev. D, 110:073005, 2024

    2024

  21. [21]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. https://arxiv.org/abs/1412.6980, 2015

    Pith/arXiv arXiv 2015

  22. [22]

    Chollet et al

    F. Chollet et al. Keras.https://keras.io. Accessed: 2026-02-02

    2026

  23. [23]

    Abadi et al

    M. Abadi et al. TensorFlow: Large-scale machine learning on heterogeneous systems.https://www. tensorflow.org/. Accessed: 2026-02-02

    2026

  24. [24]

    Blennow et al

    M. Blennow et al. Quantifying the sensitivity of oscillation experiments to the neutrino mass ordering.J. High Energ. Phys., 28(2014), 2014

    2014

  25. [25]

    Predicting the Neutrino Mass Ordering Using Neural Networks

    T.J.C. Bezerra, L. Asquith, E. Bannister, and W. Shorrock. Companion data for "Predicting the Neutrino Mass Ordering Using Neural Networks": pre-trained models and training dataset .https://doi.org/10.5281/ zenodo.20492586, June 2026. 11

    2026