Pith. sign in

REVIEW 3 major objections 4 minor 15 references

Machine Learning Tools for the IceCube-Gen2 Optical Array

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A neural network trained on a detailed ray-tracing simulation reproduces the asymmetric, wavelength-dependent photon acceptance of IceCube-Gen2's optical modules, about 200–300 times faster than the simulation.

desk verdict The OMNNSim surrogate is the one genuinely new piece and looks promising, but the effective-area comparison relies on an unstated normalization that the described architecture cannot produce; the rest is a fair status report with some overclaimed conclusions. read the letter →

arxiv 2507.07844 v1 pith:RENVR7UU submitted 2025-07-10 astro-ph.IM astro-ph.HE

classification astro-ph.IMastro-ph.HE
keywords neutrinotelescopesIceCube-Gen2opticalmodulesimulationphotonangularacceptanceneuralnetworksnormalizingflowsgraphnoisecleaning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that the multi-photomultiplier optical modules planned for IceCube-Gen2 are complex enough that the analytic approximations in the current simulation chain should be replaced by neural networks. Its chief demonstration is an optical-module simulator: a network that takes a photon's entry point, direction, and wavelength on a surface around the module and returns a detection probability for each PMT, trained on about 20 billion photons from a detailed Geant4-based simulation. The network reproduces the asymmetric and wavelength-dependent effective-area maps of the Geant4 reference better than the analytic approximation, while running roughly 200–300 times faster on a GPU. The same proceedings also report status results for two further machine-learning applications: normalizing-flow reconstruction of muon-neutrino directions and graph-neural-network noise cleaning.

What carries the argument

The load-bearing object is a two-branch neural network with a logsoftmax output over PMTs. One branch re-expresses every input in each PMT's own local coordinates and shares weights across PMTs of the same polar or equatorial type, exploiting the module's symmetry; the other branch consumes the raw absolute coordinates so the network can learn symmetry-breaking effects such as cable shadowing. The output is a normalized per-PMT detection-probability distribution, and training minimizes the Kullback–Leibler divergence between that distribution and the Geant4 reference. This combination is what lets a single network represent the full effective-area map as a function of direction and wavelength.

What would settle it

Measure the per-PMT angular and wavelength response of a real 16-PMT IceCube-Gen2 prototype module in the laboratory and compare those maps with the network's predictions. If the measured maps deviate from the Geant4-based reference by more than the analytic approximation's current errors, then the claim that the network captures the module's true photon acceptance fails even though it matches Geant4.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central result is that neural-network optical-module simulation captures both the asymmetry and wavelength dependence of the modules' photon angular acceptance, with fast GPU inference. Trained on roughly 20 billion isotropically generated photons with flat wavelength coverage from 270 nm to 700 nm, the network converts each photon's landing position, direction, and wavelength into per-PMT detection probabilities. Compared with the analytic plane-wave approximation, its 50th and 90th percentile relative errors against the Geant4 reference are lower across 3072 directions and all tested wavelengths, and it reproduces the per-PMT asymmetries that the analytic model enforces away. On a GPU it processes one million photons in 0.3 seconds, about 200–300 times faster than Geant4 on a CPU. The paper additionally reports that conditional normalizing flows give sub-degree angular resolution for starting $\nu_\mu$ charged-current events, and that a graph neural network suppresses more than 99% of radioactive noise pulses versus roughly 70% for the standard seeded cleaning.

Load-bearing premise

The network is trained on, and judged against, the same Geant4-based simulation; if that simulation gets the glass, gel pads, or PMT optical response wrong, the network is only a fast and accurate model of the simulation, not of the real module.

Editorial extensions

If this is right

  • IceCube-Gen2 event simulation can replace the analytic plane-wave approximation with network inference, making full photon-level simulation of the eight-times-larger array computationally feasible.
  • Because the network keeps wavelength as an input, the same machinery can produce wavelength-resolved effective areas for any candidate DOM design without refitting analytic functions.
  • The reconstruction results indicate that conditional normalizing flows can deliver calibrated confidence contours above 10 TeV, with sub-degree angular resolution for starting muon-neutrino events.
  • GNN-based noise cleaning removes more than 99% of radioactive-noise pulses, allowing the higher noise rates of multi-PMT modules to be handled without the physics-pulse loss of scaled SRT cuts.
  • These three pieces point toward a complete machine-learning simulation, reconstruction, and cleaning chain for the IceCube-Gen2 optical array.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the 200–300x speedup transfers to production, simulation-based inference becomes practical for IceCube-Gen2: one could generate on-the-fly training data for reconstruction, or use the network inside likelihood-free calibration loops, something the paper does not pursue.
  • Because the network outputs a full wavelength-dependent probability, it could be paired with chromatic-dispersion information in the ice to improve directional or energy estimates, an extension the paper leaves implicit.
  • The low-energy undercoverage reported for the normalizing flows suggests a direct test: augment the training set with additional sub-10 TeV events; if coverage does not approach nominal, the flow model itself, not the data balance, is the bottleneck.
  • The reliance of the optical-module network on simulated training data means its real-world validity rests on a laboratory validation campaign; without one, agreement with Geant4 is necessary but not sufficient evidence of physical accuracy.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. This ICRC proceedings paper reports three machine-learning developments for the IceCube-Gen2 optical array: (1) a neural-network surrogate for optical-module simulation, trained on OMSim/Geant4 photon-level data and compared against an analytical approximation via effective-area maps; (2) directional reconstruction of νμ charged-current events using a 3D von Mises-Fisher head and conditional normalizing flows on top of a graph-transformer backbone; and (3) a GNN-based noise-cleaning classifier compared with the classical Seeded RT algorithm. The paper claims fast GPU inference (200–300× faster than Geant4), sub-degree angular resolution for the flow-based reconstruction, and >99% noise-pulse suppression with the GNN.

Significance. The three applications are relevant and timely for IceCube-Gen2, and the paper benefits from concrete quantitative comparisons: a 200–300× speedup for optical-module simulation, coverage agreement within 5% above 10 TeV for the normalizing flows, and a large improvement in noise suppression over SRT. The authors also point to public software (OMNNSim, GraphNet, jammy_flows), which supports reproducibility. However, the central comparison of Section 2 is not reproducible as written because the network's logsoftmax output cannot represent the absolute detection probability needed for Eq. (1). Until that is clarified or fixed, the headline claim that the NN 'captures the wavelength dependence of the modules' photon angular acceptance' is not supported by the described architecture. The reconstruction and noise-cleaning sections are more self-contained, though some conclusion statements overreach the presented evidence.

major comments (3)
  1. [Sec. 2.2, Eq. (1), Figs. 1–2] The logsoftmax activation normalizes the 16 PMT outputs to sum to 1, so the network cannot represent the probability that an incoming photon is not detected by any PMT. Effective area in Eq. (1) depends on the absolute number of detected photons N_det, but the described network predicts only a conditional distribution over PMTs (or an expected number of PMT hits equal to 1 if the outputs are summed). The paper never states an additional scalar efficiency factor, a null/background class, or a target convention that preserves absolute acceptance. As a result, the Geant4-versus-NN effective-area maps in Figure 1 and the error percentiles in Figure 2 cannot be reproduced from the described architecture, and the KL-divergence training target is inconsistent if the Geant4 reference contains absolute per-PMT detection probabilities. Please specify the absolute-efficiency handling explicitly (e.g., an extra output, a post-hoc calibration factor, or a redefinition of the comparison metric) and, if such a factor exists, report its values or fitting procedure.
  2. [Sec. 3.2 and Conclusion] Figure 3 shows undercoverage at low energies and agreement with the ideal coverage only within 5% above 10 TeV. The conclusion's unqualified statement that the models provide 'reliable asymmetric uncertainty contours' overstates the presented evidence. Please qualify this claim by restricting it to the energy range where the coverage validation holds, and state whether the coverage was evaluated on an independent test set or on the training distribution.
  3. [Sec. 3 and Conclusion] The conclusion says the three techniques 'outperform the classical baselines evaluated here,' but Section 3 contains no classical reconstruction baseline: the comparison is between the 3D-vMF head and the normalizing-flow head, both of which are neural-network approaches. Either add a maximum-likelihood or other classical reconstruction comparison in Section 3, or rephrase the conclusion so that the claim is limited to the specific baselines actually shown (analytical approximation for Section 2 and SRT for Section 4).
minor comments (4)
  1. [Sec. 2.3] The inference speed of 'one million photons in 0.3 s on a GPU' is quoted without specifying the GPU model and software stack; please add this context so the 200–300× speedup is interpretable.
  2. [Sec. 3.1] The paper does not explicitly state that the coverage and angular-resolution results are computed on a validation or test sample disjoint from the 6.5-million-event training set; please state this explicitly to rule out overfitting as the source of the reported agreement.
  3. [Sec. 4] The paper states that waveform simulation is not currently performed for IceCube-Gen2 PMTs and that 'significant deviations are not expected' from mixed noise and physics hits in a single waveform. This expectation is not demonstrated; either remove the unsupported expectation or provide a concrete argument or test for it.
  4. [Sec. 2.2 and Conclusion] The claim that the network 'captures' the module's photon angular acceptance should be phrased as reproducing the OMSim/Geant4 simulation, since both training and evaluation use that same framework; this helps avoid overstating the physical accuracy of the surrogate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation chain: the NN results are supervised surrogate benchmarks, and the self-citations are methodological rather than load-bearing.

full rationale

The paper's central optical-module claim is that a neural network can approximate the OMSim/Geant4 PMT acceptance faster than the analytical PPC model. That is an empirical emulation result, not a first-principles derivation: the network is trained on Geant4/OMSim targets (Sec. 2.2) and evaluated against the same simulation in Sec. 2.3, so the agreement measures fit quality on the task it was trained for. This is the normal, non-circular logic of surrogate modeling; the paper does not claim to predict Geant4 from independent physics. The analytical approximation is a legitimate baseline, and Figure 2 quantifies errors against Geant4 rather than against the network's own training loss. Self-citations to OMNNSim [5], jammy_flows [13], and IceCube methods [12,14] are methodological: the architecture and training are described in the text (two branches, logsoftmax, KL divergence), and the reconstruction/noise-cleaning results are evaluated on independent MC samples against classical baselines (SRT, 3D-vMF). No uniqueness theorem, fitted constant, or ansatz is imported from prior work to force the present conclusions. A separate correctness/reproducibility concern exists but is not circular: the final logsoftmax normalizes the per-PMT outputs to sum to one, so the absolute detection probability needed in Eq. (1) is not specified; this is an internal gap in how A_eff maps were produced, not a derivation that reduces to its own input.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The central results rest on the fidelity of the OMSim/Geant4 reference, on summary-statistic compression of pulse information, and on several explicitly stated approximations. Free parameters include the baseline analytical fit, the hand-chosen resolution quality cut, and the scaling of the SRT baseline.

free parameters (3)
  • Analytical PMT shape parameter beta and elevation angles = Not specified; fit to Geant4 at 400 nm
    Used in the baseline analytical approximation for DOM angular acceptance; the paper notes the model may be improved with two terms per PMT type. These are fitted numbers, though they belong to the baseline and not to the neural network.
  • Quality cut sigma_flows < 1 degree = 1 degree
    Hand-chosen cut applied in Figure 4 before reporting angular resolution; it affects the reported resolution and the retention fractions, and is a post-hoc selection.
  • SRT cut scaling factors for Gen2 = Not specified
    The baseline noise-cleaning cuts were scaled proportionally to the larger inter-string spacing of Gen2; the scaling choice affects the claimed SRT noise suppression of about 70 percent.
assumptions (6)
  • domain assumption OMSim/Geant4 simulation correctly models the physical DOM response.
    The NN is trained on OMSim output and evaluated against the same framework, so the fidelity of the whole approach rests on this simulation being a faithful proxy for the real module.
  • domain assumption Pulse summary statistics preserve enough information for directional reconstruction.
    Section 3.1 follows the strategy of [9] and compresses each PMT's pulse list into summary statistics, discarding per-pulse timing correlations that could matter for reconstruction.
  • ad hoc to paper Photons that enter the enclosing sphere never leave it.
    Explicitly listed as an approximation in Section 2.2; if backscattering or escape is significant, the effective-area maps would be biased.
  • ad hoc to paper The ice inside the enclosing sphere is modeled only by its wavelength-dependent refractive index.
    Explicitly listed as an approximation in Section 2.2; absorption and scattering inside the sphere are ignored.
  • domain assumption The GNN noise-cleaning model trained on events with at most 1e5 pulses generalizes to higher pulse counts.
    Section 4 states the model was trained only on events with up to 1e5 pulses due to computational resources; the performance curves extend to 1e5.
  • ad hoc to paper Mixed noise and physics hits within a single waveform will not significantly change the noise-cleaning results.
    Section 4 notes waveform simulation is not currently performed for IceCube-Gen2 PMTs and significant deviations are not expected, which is an unverified assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Learning Tools for the IceCube-Gen2 Optical Array." pith.science (2026). https://pith.science/paper/RENVR7UU

@misc{pith2026250707844,
  author       = {Pith},
  title        = {Pith review of: Machine Learning Tools for the IceCube-Gen2 Optical Array},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RENVR7UU}},
  note         = {Machine review of arXiv:2507.07844}
}
read the original abstract

Neural networks (NNs) have a great potential for future neutrino telescopes such as IceCube-Gen2, the planned high-energy extension of the IceCube observatory. IceCube-Gen2 will feature new optical sensors with multiple photomultiplier tubes (PMTs) designed to provide omnidirectional sensitivity. Neural networks excel at handling high-dimensional problems and can naturally incorporate the increased complexity of these new sensors. Additionally, their fast inference time makes them promising candidates for handling the high event rates expected from IceCube-Gen2. This contribution presents potential applications of neural networks in the IceCube-Gen2 in-ice optical array. First, we introduce a method to simulate the IceCube-Gen2 optical modules' photon acceptance using a NN that leverages the modules' inherent symmetries. Secondly, we present the status of neutrino NN-based reconstruction efforts, including the adaptation of a novel IceCube technique that combines normalizing flows with transformer NNs. Finally, we describe current progress in noise cleaning applications based on node classification with graph neural networks (GNNs), a method that has already shown promising results for the forthcoming low-energy extension, IceCube-Upgrade.

Figures

Figures reproduced from arXiv: 2507.07844 by the authors.

Figure 1
Figure 1. Omnidirectional maps of effective area 𝐴eff for a single equatorial PMT as a function of incident direction. From left to right: Geant4, Neural Network, and Analytical Approximation: (a) 400 nm, (b) 550 nm. 300 400 500 600 Wavelength (nm) 10 1 10 0 10 1 | Aeff|/A true eff (%) IceCube-Gen2 Preliminary Full Optical Module 300 400 500 600 Wavelength (nm) 10 0 10 1 10 2 IceCube-Gen2 Preliminary Polar PMT 300 400 500 600… view at source ↗
Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Predicted vs. observed coverage for conditional normalizing flows for different deposited energies. 3.2 Results One of the most interesting aspects of conditional normalizing flows is their ability to predict asymmetric confidence contours [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Angular resolution as a function of deposited energy for the different approaches for starting and through-going 𝜈𝜇 CC. We followed a similar approach to that of [14] and trained a graph neural network to perform the cleaning. Our baseline is an adaptation of the DynEd…
Figure 5
Figure 5. Figure 5: Remaining noise (a) and signal (b) pulse fractions versus total pulses per event, comparing the GNN-based cleaning to classical SRT; vertical purple lines in (b) mark the mean expected noise level ± 1 𝜎. DOM glass, compared with about 70 % for the classical SRT algorit…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 12 canonical work pages

  1. [1]

    Kowalski,PoSICRC2025(these proceedings) 1080

    IceCube-Gen2 Collaboration, M. Kowalski,PoSICRC2025(these proceedings) 1080

  2. [2]

    Kappes,PoSICRC2025(these proceedings) 1072

    IceCube-Gen2 Collaboration, A. Kappes,PoSICRC2025(these proceedings) 1072

  3. [3]

    Chirkin,Nucl

    IceCubeCollaboration, D. Chirkin,Nucl. Instrum. Meth. A 725 (2013) 141–143

  4. [4]

    IceCube Collaboration, “OMSim.”https://github.com/icecube/OMSim. 2025

  5. [5]

    OMNNSim

    F. J. Vara Carbonell, “OMNNSim.”https://github.com/jvaracarbonell/OMNNSim. 2025

  6. [6]

    Kullback and R

    S. Kullback and R. A. Leibler,Ann. Math. Stat. 22no. 1, (Mar., 1951) 79–86

  7. [7]

    Schatto,Stacked searches for high-energy neutrinos from blazars with IceCube

    K. Schatto,Stacked searches for high-energy neutrinos from blazars with IceCube. PhD thesis, 2014

  8. [8]

    Søgaardet al., JOSS 8no

    A. Søgaardet al., JOSS 8no. 85, (May, 2023) 4971

Show all 15 references
  1. [9]

    and others,JINST16 no

    IceCubeCollaboration, Abbasi, R. and others,JINST16 no. 07, (2021) P07041

  2. [10]

    Icecube – neutrinos in deep ice the top 3 solutions from the public kaggle competition,

    H. Bukhariet al., “Icecube – neutrinos in deep ice the top 3 solutions from the public kaggle competition,” 2023.https://arxiv.org/abs/2310.15674

  3. [11]

    Rethinking attention with performers,

    K. Choromanskiet al., “Rethinking attention with performers,” 2022. https://arxiv.org/abs/2009.14794

  4. [12]

    Glüsenkamp,PoSICRC2023(2023) 1003

    IceCubeCollaboration, T. Glüsenkamp,PoSICRC2023(2023) 1003

  5. [13]

    jammy_flows

    T. Glüsenkamp, “jammy_flows.”https://github.com/thoglu/jammy_flows. 2022

  6. [14]

    Elleret al., PoSICRC2023(2023) 1036

    IceCubeCollaboration, P. Elleret al., PoSICRC2023(2023) 1036

  7. [15]

    Abbasiet al., JINST17no

    IceCubeCollaboration, R. Abbasiet al., JINST17no. 11, (2022) P11003. 8 Machine Learning Tools for the IceCube-Gen2 Optical Array Full Author List: IceCube-Gen2 Collaboration R. Abbasi16, M. Ackermann76, J. Adams21, S. K. Agarwalla46, a, J. A. Aguilar10, M. Ahlers25, J.M. Alame...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.