{"id":"9e1f8907-050f-4499-89ca-3963fa62ddb1","arxiv_id":"2507.07844","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Neural-network surrogates for optical-module response, normalizing-flow neutrino reconstruction, and GNN noise cleaning all show promise for IceCube-Gen2, though the reconstruction undercovers low-energy events.","lead":"This proceedings paper reports three machine-learning tools being developed for the IceCube-Gen2 neutrino telescope: a neural network that simulates the new optical modules' photon response, transformer-based neutrino direction reconstruction, and graph-neural-network noise cleaning. The results are preliminary but show the ML tools matching or beating the classical methods used for the current IceCube detector.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The logsoftmax output normalizes per-PMT probabilities to sum 1, so undetected photons cannot be represented; the reported effective-area maps require an unstated absolute-efficiency factor, leaving the central comparison non-reproducible as written.","rationale":"The reader's weakest assumption concerns external validity: the network is trained and evaluated on OMSim/Geant4, so if that simulation is physically wrong, the NN cannot be more accurate than its training data. That concern is real but does not affect the internal claim that the NN reproduces Geant4 better than the analytical approximation. I found a more immediate, load-bearing gap: as described, the network's logsoftmax output cannot represent undetected photons, so the absolute effective-area comparisons in Figures 1 and 2 require an unstated normalization step. Without knowing how absolute detection efficiency enters, the central quantitative evidence for the paper's headline claim is not reproducible even under the assumption that OMSim is perfect. This is an internal consistency issue, not merely a disagreement with external calibration. The verdict should remain CONDITIONAL: the paper needs a precise description of the normalization/efficiency mechanism and a verification that Figure 2 uses it, before the claim can be accepted as stated. If the mechanism turns out to be absent, the effective-area results would be invalid and the verdict should be REJECT; the concrete test distinguishes these cases.","tokens_in":11633,"tokens_out":8228,"duration_ms":99455,"concrete_test":"Inspect the OMNNSim inference code (github.com/jvaracarbonell/OMNNSim) and rerun the Figure 2 evaluation. For a test set of photons with known Geant4/OMSim labels, record the average sum of the 16 PMT output probabilities separately for photons that OMSim labels as detected and undetected. (1) If the average sum is 1 for undetected photons, identify the external normalization/efficiency factor used to produce Figure 1 and confirm it is applied consistently across wavelength and direction; recompute Figure 2 with that factor explicitly included. (2) If the model instead has a 17th 'no-detection' class or a separate efficiency head, document this and verify that the absolute effective areas in Figure 1 match the network's actual output.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Section 2.2, the network 'outputs the detection probability for each PMT' but then applies a final logsoftmax activation so the outputs are normalized and 'can be interpreted as a PDF.' This means the 16 PMT probabilities sum to 1 for every incoming photon, leaving no way to represent the case where a photon enters the enclosing sphere but is not detected by any PMT. The effective area in Eq. (1) depends on the absolute number of detected photons, N_det, so the Figure 1 and Figure 2 comparisons require either an additional scalar efficiency factor, a null/background class inside the softmax, or a target convention that preserves absolute acceptance. None of these is stated in the paper. If the Geant4/OMSim targets are absolute per-PMT detection probabilities, a KL divergence against a logsoftmax-normalized output would not learn the overall detection probability; it would learn only the conditional distribution over PMTs given a detection. If the network can only predict conditional assignments, then the 24.89 and 11.86 cm^2 effective-area values in Figure 1 cannot be produced by the network alone, and the claim that the NN 'captures the wavelength dependence of the modules' photon angular acceptance' is not supported by the described architecture. This is an internal reproducibility gap, independent of whether OMSim is an accurate model of the real DOM.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":11934,"tokens_out":4724,"duration_ms":57578,"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":[{"comment":"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.","section":"Sec. 2.2, Eq. (1), Figs. 1–2"},{"comment":"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.","section":"Sec. 3.2 and Conclusion"},{"comment":"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).","section":"Sec. 3 and Conclusion"}],"minor_comments":[{"comment":"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.","section":"Sec. 2.3"},{"comment":"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.","section":"Sec. 3.1"},{"comment":"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.","section":"Sec. 4"},{"comment":"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.","section":"Sec. 2.2 and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The logsoftmax issue in Section 2 is the main technical blocker. If the authors can clarify that absolute effective areas are obtained by multiplying the normalized output by a separately determined efficiency (and report that factor), the result may be sound; as written, the central comparison is internally inconsistent. The reconstruction and noise-cleaning sections are reasonable for a proceedings contribution, but the conclusion should be aligned with the shown energy range and baseline comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a well-written ICRC proceedings from IceCube-Gen2, and the one genuinely new result is OMNNSim—a symmetry-aware neural surrogate for the optical module's photon acceptance. It beats the current analytical approximation on Geant4 effective-area maps, captures the asymmetries and wavelength dependence, and runs 200–300x faster on GPU. That is a real tool, and the collaboration deserves credit for putting it forward with comparisons.\n\nThe other two sections are adaptations of known IceCube techniques to Gen2: normalizing-flow reconstruction and GNN noise cleaning. The evaluations look solid: coverage within 5% above 10 TeV, and >99% noise suppression vs roughly 70% for SRT. The paper is transparent that this is preliminary, which helps.\n\nThe soft spots are in the presentation of the OMNNSim result. The network's final logsoftmax forces the 16 PMT outputs to sum to 1. That leaves no representable outcome for a photon that enters the sphere but is not detected by any PMT. Effective area in Eq. (1) depends on the absolute number of detected photons. The paper never says how the normalization to absolute acceptance is done—a separate efficiency factor, a null class, or a target convention. Without that, the maps in Figures 1 and 2 cannot be reproduced from the described architecture. This is an internal gap, independent of whether OMSim is a good model of the real DOM. The paper's claim that the NN captures the angular acceptance would be correct for the conditional per-PMT distribution, but not for the absolute effective area.\n\nAlso, the conclusion calls the uncertainty contours reliable, but the coverage plot shows undercoverage below 10 TeV. That is an overstatement; it should say reliable above 10 TeV. And the sub-degree angular resolution headline depends on the sigma_flows < 1 degree cut that retains only 40–65% of events; that caveat needs to travel with the number. Minor: the noise-cleaning comparison has no uncertainties on the 99%/70% numbers, and the SRT scaling is a reasonable but stated approximation. These are fixable in a final version.\n\nWho is this for? Practitioners in neutrino astronomy and anyone benchmarking ML for detector simulation. It deserves a serious referee—the OMNNSim result is new and the evaluations are useful—but the normalization issue should be resolved or explicitly documented before the central claim is taken as established.","headline":"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.","tokens_in":12463,"tokens_out":3993,"would_cite":true,"duration_ms":43267,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["neutrino telescopes","IceCube-Gen2","optical module simulation","photon angular acceptance","neural networks","normalizing flows","graph neural networks","noise cleaning"],"falsifier":"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.","tokens_in":11394,"feed_emoji":"🔭","tokens_out":9928,"duration_ms":98059,"temperature":0.7,"pith_summary":"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.","feed_headline":"Neural net reproduces IceCube-Gen2 optics 300x faster","feed_subtitle":"Trained on a detailed simulation, the network captures angular asymmetries and wavelength dependence that analytic fits miss.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Defines IceCube-Gen2 and the scale of the planned array, motivating the need for fast simulation and reconstruction.","marker":"[1]"},{"why":"Supplies the 16-PMT prototype DOM design used as the baseline module for the simulation study.","marker":"[2]"},{"why":"Provides the PPC plane-wave analytical approximation that the neural network is compared against.","marker":"[3]"},{"why":"Supplies the Geant4-based optical simulation chain that generated the roughly 20 billion training photons.","marker":"[4]"},{"why":"Provides the neural-network implementation used to train and evaluate the optical-module simulator.","marker":"[5]"},{"why":"Defines the Kullback–Leibler divergence objective that aligns the network's output PDF with the Geant4 reference.","marker":"[6]"}],"fun_headline_variants":["Neural network speeds IceCube-Gen2 optics simulation 300x","AI reproduces IceCube-Gen2 photon acceptance 300x faster","Machine learning captures IceCube-Gen2 optical asymmetries","ML tools advance IceCube-Gen2 simulation and cleaning","Deep learning boosts IceCube-Gen2 reconstruction and noise filtering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Neural network speeds IceCube-Gen2 optics simulation 300x","AI reproduces IceCube-Gen2 photon acceptance 300x faster","Machine learning captures IceCube-Gen2 optical asymmetries","ML tools advance IceCube-Gen2 simulation and cleaning","Deep learning boosts IceCube-Gen2 reconstruction and noise filtering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1592,"prompt_tokens":968,"completion_tokens":624,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":536}},"tokens_in":584,"tokens_out":624,"duration_ms":6330,"temperature":1.0,"reasoning_tokens":536,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:30:39.118259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Kowalski,PoSICRC2025(these proceedings) 1080","cited_arxiv_id":null,"evidence_quote":"Defines IceCube-Gen2 and the scale of the planned array, motivating the need for fast simulation and reconstruction."},{"cited_title":"Kappes,PoSICRC2025(these proceedings) 1072","cited_arxiv_id":null,"evidence_quote":"Supplies the 16-PMT prototype DOM design used as the baseline module for the simulation study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Geant4-based optical simulation chain that generated the roughly 20 billion training photons."},{"cited_title":"OMNNSim","cited_arxiv_id":null,"evidence_quote":"Provides the neural-network implementation used to train and evaluate the optical-module simulator."},{"cited_title":"Kullback and R","cited_arxiv_id":null,"evidence_quote":"Defines the Kullback–Leibler divergence objective that aligns the network's output PDF with the Geant4 reference."}],"review_version":1}