pith. sign in

arxiv: 2511.03910 · v1 · submitted 2025-11-05 · 🌌 astro-ph.IM · astro-ph.HE· hep-ex

Event Reconstruction for Radio-Based In-Ice Neutrino Detectors with Neural Posterior Estimation

Nils Heyer , Christian Glaser , Thorsten Gl\"usenkamp , Martin Ravn This is my paper

Pith reviewed 2026-05-18 00:29 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.HEhep-ex
keywords neutrino reconstructionradio detectionin-ice detectorsneural posterior estimationnormalizing flowsultra-high-energy neutrinosevent topologyposterior PDF
0
0 comments X

The pith

A neural network with conditional normalizing flows reconstructs neutrino energy, direction and flavor from radio waveforms, predicting full posterior distributions for uncertainties.

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

The paper develops a deep neural network to extract neutrino direction, shower energy and interaction topology directly from raw radio antenna waveforms in in-ice detectors. It applies neural posterior estimation with conditional normalizing flows to output the complete probability distribution over energy and direction for each event, rather than single point estimates. This yields improved median resolutions of 0.30 in log(E) and 18 square degrees for shallow components and 0.08 in log(E) and 28 square degrees for deep components on neutral-current events at 1 EeV. The method also reconstructs the more variable charged-current electron-neutrino events and supplies a goodness-of-fit score to check whether measured signals are consistent with the Monte Carlo training set. A sympathetic reader would care because event-by-event uncertainties and better handling of stochastic topologies could tighten limits on ultra-high-energy neutrino fluxes and sources.

Core claim

A deep neural network trained on Monte Carlo simulations reconstructs the neutrino direction, the energy of the induced particle shower, and the event topology from raw radio waveforms. For the first time the network outputs the full posterior probability density function for energy and direction by means of conditional normalizing flows, which directly supplies per-event uncertainty estimates. On neutral-current events at a shower energy of 1 EeV the approach achieves a median resolution of 0.30 in log(E) and 18 square degrees for a shallow detector component and 0.08 in log(E) and 28 square degrees for a deep component, outperforming earlier reconstruction algorithms while also handling st

What carries the argument

Conditional normalizing flows inside a neural posterior estimation network that model the full posterior distribution of neutrino parameters conditioned on the recorded radio waveforms.

If this is right

  • Event-by-event uncertainty estimates become available for downstream statistical analyses of ultra-high-energy neutrino fluxes.
  • Reconstruction extends to the more stochastic charged-current electron-neutrino events that were previously difficult to handle.
  • The impact of different antenna types and systematic uncertainties on resolution can be quantified directly from the network output.
  • A goodness-of-fit score derived from the posterior allows rejection of events whose waveforms are incompatible with the training simulations.

Where Pith is reading between the lines

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

  • The same architecture could be retrained on hybrid optical-radio data sets to cross-calibrate energy scales between detector technologies.
  • If the posterior widths prove reliable, they could be used to weight events in source-association studies without additional simulation campaigns.
  • Detector design studies could replace slow template-fitting reconstructions with this fast network to scan larger parameter spaces of antenna spacing and depth.

Load-bearing premise

The Monte Carlo simulations used for training accurately capture all relevant detector responses, neutrino interaction physics, and systematic effects so that the trained model generalizes to real measured signals.

What would settle it

A large, statistically significant discrepancy between the posterior distributions predicted by the network on real data and the distributions expected from independent Monte Carlo simulations of the same detector configuration.

read the original abstract

The detection of ultra-high-energy (UHE) neutrinos in the EeV range is the goal of current and future in-ice radio arrays at the South Pole and in Greenland. Here, we present a deep neural network that can reconstruct the main neutrino properties of interest from the raw waveforms recorded by the radio antennas: the neutrino direction, the energy of the particle shower induced by the neutrino interaction, and the event topology, thereby estimating the neutrino flavor. For the first time, we predict the full posterior PDF for the energy and direction reconstruction via neural posterior estimation utilizing conditional normalizing flows, enabling event-by-event uncertainty prediction. We improve over previous reconstruction algorithms and obtain a median resolution of 0.30 log(E) and 18 square degrees for a 'shallow' detector component and 0.08 log(E) and 28 square degrees for a 'deep' detector component for neutral current (NC) events at a shower energy of 1 EeV. This deep learning approach also allows us to reconstruct the more stochastic $\nu_e$ - charged current (CC) events. We quantify the impact of different antenna types and systematic uncertainties on the reconstruction and derive a goodness-of-fit score to test the compatibility of measured neutrino signals with the Monte Carlo simulations used to train the neural network.

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 / 2 minor

Summary. The manuscript presents a deep neural network using conditional normalizing flows for neural posterior estimation (NPE) to reconstruct neutrino direction, energy, and event topology (including flavor via CC/NC classification) from raw radio waveforms in in-ice detectors. It reports improved median resolutions over prior algorithms—0.30 log(E) and 18 deg² (shallow) and 0.08 log(E) and 28 deg² (deep) for 1 EeV NC events—along with full posterior PDFs for event-by-event uncertainties, quantification of antenna and systematic effects, and a goodness-of-fit score for MC compatibility.

Significance. If the MC-derived posteriors and resolutions generalize, the work would advance reconstruction for UHE radio neutrino arrays by enabling calibrated uncertainty estimates and handling of stochastic ν_e CC events, strengthening sensitivity projections for experiments at the South Pole and Greenland. The explicit use of normalizing flows for full posteriors and the GoF metric are positive steps toward reproducible, uncertainty-aware analysis.

major comments (2)
  1. [Abstract and results section] Abstract and results section: the headline resolutions (0.30 log(E), 18 deg² shallow; 0.08 log(E), 28 deg² deep for 1 EeV NC) and the claim of calibrated event-by-event posteriors are obtained exclusively on the Monte Carlo training distribution; the introduced goodness-of-fit score tests compatibility with the same simulations but does not directly probe calibration under unmodeled domain shifts in detector response, ice properties, or noise realizations outside the spanned variations.
  2. [Validation and systematics discussion] Validation and systematics discussion: while selected systematics are quantified, no independent hold-out dataset, real-data proxy, or stress test of posterior coverage under realistic mismatches is presented, leaving the load-bearing assumption that the learned conditional density matches the true data-generating process unverified for deployment on measured signals.
minor comments (2)
  1. [Introduction] Define 'shallow' and 'deep' detector components explicitly in the introduction or methods before quoting component-specific resolutions.
  2. [Methods] Add explicit details on training/validation splits, hyperparameter selection, and how uncertainties are propagated from the flow to the reported median resolutions.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their constructive and detailed report. We address each major comment below, indicating where revisions have been made to the manuscript.

read point-by-point responses

  1. Referee: [Abstract and results section] Abstract and results section: the headline resolutions (0.30 log(E), 18 deg² shallow; 0.08 log(E), 28 deg² deep for 1 EeV NC) and the claim of calibrated event-by-event posteriors are obtained exclusively on the Monte Carlo training distribution; the introduced goodness-of-fit score tests compatibility with the same simulations but does not directly probe calibration under unmodeled domain shifts in detector response, ice properties, or noise realizations outside the spanned variations.

    Authors: We agree that the headline resolutions and posterior calibration results are evaluated on the Monte Carlo training distribution. This is a standard limitation in the field, as no real UHE neutrino events have been recorded by these detectors to date. The goodness-of-fit score is explicitly designed to test per-event compatibility with the training simulations and can flag potential mismatches. In the revised manuscript we have expanded the discussion of these limitations, including the assumptions about domain shifts and the role of the GoF metric in future analyses with real data. revision: partial

  2. Referee: [Validation and systematics discussion] Validation and systematics discussion: while selected systematics are quantified, no independent hold-out dataset, real-data proxy, or stress test of posterior coverage under realistic mismatches is presented, leaving the load-bearing assumption that the learned conditional density matches the true data-generating process unverified for deployment on measured signals.

    Authors: We acknowledge that the manuscript does not include an independent real-data hold-out set or explicit stress tests for unmodeled mismatches outside the simulated variations. Selected systematics (antenna types, ice properties) are quantified within the ranges covered by the training simulations. In the revision we will add controlled stress tests that perturb noise realizations and ice parameters beyond the training distribution to evaluate posterior coverage under such mismatches, thereby providing a clearer assessment of robustness. revision: yes

standing simulated objections not resolved
  • Direct validation of posterior calibration and resolution on actual measured UHE neutrino signals, as no such events have been observed with in-ice radio detectors to date.

Circularity Check

1 steps flagged

Resolution claims remain internal to Monte Carlo training distribution

specific steps
  1. fitted input called prediction [Abstract]
    "We improve over previous reconstruction algorithms and obtain a median resolution of 0.30 log(E) and 18 square degrees for a 'shallow' detector component and 0.08 log(E) and 28 square degrees for a 'deep' detector component for neutral current (NC) events at a shower energy of 1 EeV."

    The quoted median resolutions for energy and direction are computed by running the trained NPE model on Monte Carlo events drawn from the same simulation framework and parameter variations used to generate the training set. This renders the reported performance numbers a statistical summary of how well the learned posterior matches the training distribution rather than an out-of-sample or real-data prediction.

full rationale

The paper trains a conditional normalizing flow model for neural posterior estimation exclusively on Monte Carlo simulations of in-ice neutrino interactions. The central performance claims (median resolutions of 0.30 log(E) and 18 deg² shallow / 0.08 log(E) and 28 deg² deep at 1 EeV for NC events) are obtained by evaluating the same model on held-out events from that identical simulation ensemble. A goodness-of-fit score is defined to test compatibility with the training simulations, but this does not convert the quoted resolutions into an independent prediction; they remain a direct measure of in-distribution reconstruction fidelity. No equations reduce by algebraic construction and no load-bearing self-citation chain is present, so the circularity is moderate rather than definitional.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The reconstruction performance rests on the fidelity of the Monte Carlo training data and the assumption that the neural network generalizes beyond the simulated distributions.

free parameters (1)
  • Neural network parameters and flow hyperparameters
    Weights and architecture choices are optimized on simulated data to achieve the quoted resolutions.
axioms (1)
  • domain assumption Monte Carlo simulations of radio signals and neutrino interactions match real detector behavior sufficiently for generalization
    All training and reported performance metrics depend on this match.

pith-pipeline@v0.9.0 · 5776 in / 1312 out tokens · 42473 ms · 2026-05-18T00:29:14.225782+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

46 extracted references · 46 canonical work pages

  1. [1]

    IceCube Collaboration,Evidence for High-Energy Extraterrestrial Neutrinos at the IceCube Detector, Science342(2013) 1242856

    work page 2013

  2. [2]

    IceCube Collaboration,Neutrino emission from the direction of the blazar TXS 0506+056 prior to the IceCube-170922A alert,Science361(2018) 147–151

    work page 2018

  3. [3]

    IceCube Collaboration,Evidence for neutrino emission from the nearby active galaxy NGC 1068,Science378 (2022) 538–543

    work page 2022

  4. [4]

    IceCube Collaboration,Observation of high-energy neutrinos from the Galactic plane,Science380(2023) 1338–1343

    work page 2023

  5. [5]

    KM3NeT Collaboration,Observation of an ultra-high-energy cosmic neutrino with km3net,Nature 638(2025) 376–382

    work page 2025

  6. [6]

    Barwick and C

    S. Barwick and C. Glaser,Radio Detection of High Energy Neutrinos in Ice,The Encyclopedia of Cosmology2(2023) 237–302, [arXiv:2208.04971]

    work page arXiv 2023

  7. [7]

    V. B. Valera, M. Bustamante and C. Glaser,Near-future discovery of the diffuse flux of ultrahigh-energy cosmic neutrinos,Phys. Rev. D107(2023) 043019

    work page 2023

  8. [8]

    D. F. Fiorillo, M. Bustamante and V. B. Valera, Near-future discovery of point sources of ultra-high-energy neutrinos,J. Cosmol. Astropart. Phys.03(2023) 026

    work page 2023

  9. [9]

    V. B. Valera, M. Bustamante and C. Glaser,The ultra-high-energy neutrino-nucleon cross section: measurement forecasts for an era of cosmic EeV-neutrino discovery,J. High Energ. Phys.06 (2022) 105

    work page 2022

  10. [10]

    Coleman, O

    A. Coleman, O. Ericsson, C. Glaser and M. Bustamante,Flavor composition of ultrahigh-energy cosmic neutrinos: Measurement forecasts for in-ice radio-based EeV neutrino telescopes,Phys. Rev. D110 (2024) 023044. 19

    work page 2024

  11. [11]

    G. A. Askar’yan,Coherent Radio Emission from Cosmic Showers in Air and in Dense Media,Soviet Physics JETP-USSR21(1965) 658

    work page 1965

  12. [12]

    Saltzberg, P

    D. Saltzberg, P. Gorham, D. Walz, C. Field, R. Iverson et al.,Observation of the Askaryan Effect: Coherent Microwave Cherenkov Emission from Charge Asymmetry in High-Energy Particle Cascades,Phys. Rev. Lett.86(2001) 2802–2805

    work page 2001

  13. [13]

    P. W. Gorham, D. Saltzberg, R. C. Field, E. Guillian, R. Milinˇ ci´ c et al.,Accelerator measurements of the Askaryan effect in rock salt: A roadmap toward teraton underground neutrino detectors,Phys. Rev. D72 (2005) 023002

    work page 2005

  14. [14]

    ANITA Collaboration,Observations of the askaryan effect in ice,Phys. Rev. Lett.99(2007) 171101

    work page 2007

  15. [15]

    Barwick, D

    S. Barwick, D. Besson, P. Gorham and D. Saltzberg, South Polar in situ radio-frequency ice attenuation, Journal of Glaciology51(2005) 231–238

    work page 2005

  16. [16]

    J. Avva, J. M. Kovac, C. Miki, D. Saltzberg and A. G. Vieregg,An in situ measurement of the radio-frequency attenuation in ice at summit station, greenland, Journal of Glaciology61(2015) 1005–1011

    work page 2015

  17. [17]

    M. G. Aartsen, R. Abbasi, M. Ackermann, J. Adams, J. A. Aguilar et al.,IceCube-Gen2: the window to tPhysRevD.110.023044he extreme Universe,J. Phys. G: Nucl. Part. Phys.48(2021) 060501

    work page 2021

  18. [18]

    IceCube-Gen2 Collaboration,IceCube-Gen2 Technical Design Report, IceCube-Gen2 website (2023)

    work page 2023

  19. [19]

    RNO-G collaboration,Design and sensitivity of the radio neutrino observatory in greenland (rno-g), Journal of Instrumentation16(2021) P03025

    work page 2021

  20. [20]

    D. J. Rezende and S. Mohamed,Variational Inference with Normalizing Flows,Int. conference on machine learning(2015) 1530–1538

    work page 2015

  21. [21]

    Plaisier, S

    I. Plaisier, S. Bouma and A. Nelles,Reconstructing the arrival direction of neutrinos in deep in-ice radio detectors,Eur. Phys. J. C83(2023) 443

    work page 2023

  22. [22]

    Anker, S

    A. Anker, S. Barwick, H. Bernhoff, D. Besson, N. Bingefors et al.,Neutrino vertex reconstruction with in-ice radio detectors using surface reflections and implications for the neutrino energy resolution,J. Cosmol. Astropart. Phys.11(2019) 030

    work page 2019

  23. [23]

    Glaser, D

    C. Glaser, D. Garc´ ıa-Fern´ andez, A. Nelles, J. Alvarez-Mu˜ niz, S. W. Barwick et al.,NuRadioMC: simulating the radio emission of neutrinos from interaction to detector,Eur. Phys. J. C80(2020) 77

    work page 2020

  24. [24]

    Barwick, E

    S. Barwick, E. Berg, D. Besson, G. Gaswint, C. Glaser et al.,Observation of classically ’forbidden’ electromagnetic wave propagation and implications for neutrino detection.,J. Cosmol. Astropart. Phys.07 (2018) 055

    work page 2018

  25. [25]

    L. D. Landau and I. Pomeranchuk,Limits of applicability of the theory of bremsstrahlung electrons and pair production at high-energies,Dokl. Akad. Nauk Ser. Fiz.92(1953) 535–536

    work page 1953

  26. [26]

    A. B. Migdal,Bremsstrahlung and pair production in condensed media at high energies,Phys. Rev.103 (1956) 1811–1820

    work page 1956

  27. [27]

    Alvarez-Mu˜ niz, R

    J. Alvarez-Mu˜ niz, R. A. V´ azquez and E. Zas, Characterization of neutrino signals with radiopulses in dense media through the Landau-Pomeranchuk-Migdal effect,Phys. Rev. D61(1999) 023001

    work page 1999

  28. [28]

    Glaser, A

    C. Glaser, A. Nelles, I. Plaisier, C. Welling, S. W. Barwick et al.,NuRadioReco: a reconstruction framework for radio neutrino detectors,Eur. Phys. J. C 79(2019) 464

    work page 2019

  29. [29]

    ARIANNA Collaboration,Capabilities of ARIANNA: Neutrino Pointing Resolution and Implications for Future Ultra-high Energy Neutrino Astronomy, Proceedings of Science: ICRC(2021) 1151

    work page 2021

  30. [30]

    RNO-G collaboration,Reconstructing the neutrino energy for in-ice radio detectors,Eur. Phys. J. C82 (2022) 147

    work page 2022

  31. [31]

    IceCube-Gen2 Collaboration,Direction reconstruction performance for IceCube-Gen2 Radio,Proceedings of Science: ICRC(2023) 1045

    work page 2023

  32. [32]

    Pierre Auger Collaboration,Measurement of the depth of maximum of air-shower profiles with energies between10 18.5 and10 20 eVusing the surface detector of the Pierre Auger Observatory and deep learning, Phys. Rev. D111(2025) 022003

    work page 2025

  33. [33]

    Pierre Auger Collaboration,Inference of the Mass Composition of Cosmic Rays with Energies from10 18.5 to10 20 eVUsing the Pierre Auger Observatory and Deep Learning,Phys. Rev. Lett.134(2025) 021001

    work page 2025

  34. [34]

    Glaser, S

    C. Glaser, S. McAleer, S. Stj¨ arnholm, P. Baldi and S. Barwick,Deep-learning-based reconstruction of the neutrino direction and energy for in-ice radio detectors, Astropart. Phys.145(2023) 102781

    work page 2023

  35. [35]

    ARA Collaboration,A neural network based UHE neutrino reconstruction method for the Askaryan Radio Array (ARA),Proceedings of Science: ICRC(2021) 1157

    work page 2021

  36. [36]

    S. W. Barwick,ARIANNA: A New Concept for UHE Neutrino Detection,Journal of Physics: Conference Series60(2007) 276

    work page 2007

  37. [37]

    Allison, S

    P. Allison, S. Archambault, R. Bard, J. Beatty, M. Beheler-Amass et al.,Design and performance of an interferometric trigger array for radio detection of high-energy neutrinos,Nucl. Instrum. Methods Phys. Res. A930(2019) 112–125

    work page 2019

  38. [38]

    Alvarez-Mu˜ niz, P

    J. Alvarez-Mu˜ niz, P. M. Hansen, A. Romero-Wolf and E. Zas,Askaryan radiation from neutrino-induced showers in ice,Phys. Rev. D101(2020) 083005

    work page 2020

  39. [39]

    K. He, X. Zhang, S. Ren and J. Sun,Deep Residual Learning for Image Recognition,IEEE Conference on Computer Vision and Pattern Recognition(2016) 770–778

    work page 2016

  40. [40]

    Seferbekov and D

    S. Seferbekov and D. Kanonik. Kaggle challange: G2Net Gravitational Wave Detection, (2021)

    work page 2021

  41. [41]

    Gl¨ usenkamp,Unifying supervised learning and VAEs: coverage, systematics and goodness-of-fit in normalizing-flow based neural network models for astro-particle reconstructions,Eur

    T. Gl¨ usenkamp,Unifying supervised learning and VAEs: coverage, systematics and goodness-of-fit in normalizing-flow based neural network models for astro-particle reconstructions,Eur. Phys. J. C84 (2024) 163

    work page 2024

  42. [42]

    Gl¨ usenkamp

    T. Gl¨ usenkamp. GitHub: jammyflows, (2024)

    work page 2024

  43. [43]

    C. Meng, Y. Song, J. Song and S. Ermon, Gaussianization Flows,Proceedings of Machine Learning Research108(2020) 4336–4345

    work page 2020

  44. [44]

    Kullback and R

    S. Kullback and R. A. Leibler,On information and sufficiency,The Annals of Mathematical Statistics22 (1951) 79–86

    work page 1951

  45. [45]

    IceCube Collaboration,Efficient propagation of systematic uncertainties from calibration to analysis with the SnowStorm method in IceCube,J. Cosmol. Astropart. Phys.2019(2019) 048

    work page 2019

  46. [46]

    M. Ravn, C. Glaser, T. Gl¨ usenkamp, A. ¨Ocelikkale and A. Coleman,Likelihood Reconstruction for Radio Detectors of Neutrinos and Cosmic Rays, arXiv:2510.21925. 20 Fig. 13Coverage for the energy and direction reconstruction for the ’shallow’ and ’deep’ components. The data was split by event topology withν x - NC in blue andν e - CC in red. Appendix A: Co...

    work page arXiv