Observation of high-energy neutrinos from the Galactic plane

• Cascades are short-ranged showers, mainly from electron- and tau-neutrino interactions and scattering by all three flavors, appearing almost point-like at IceCube’s spacing.Source: Cascade events in IceCube · p. 3
• Atmospheric muons dominate IceCube’s background at about 100 million per observed astrophysical neutrino, while track-based searches are most sensitive to the Northern sky.The Galactic Center and most expected Galactic-plane neutrino emission lie in the Southern sky, where selecting interactions observed within the detector greatly reduces entering cosmic-ray muons.Source: Cascade events in IceCube · pp. 3–4
• At TeV energies, selecting cascades reduces atmospheric-neutrino contamination by about an order of magnitude and lowers the analysis threshold to about 1 TeV.Source: Cascade events in IceCube · p. 4
• In the Southern sky, cascades’ lower background, better energy resolution, and lower threshold compensate for inferior angular resolution, benefiting searches for extended emission.Cascade analyses rely more on signal purity than individual-event angular resolution, and are expected to be more sensitive to extended TeV emission from the Southern sky.Source: Cascade events in IceCube · pp. 4–5

• CNN-based selection (15,28) runs in milliseconds per event, enabling complex early filtering that retains lower-energy neutrinos and difficult boundary events.Source: Application of deep learning to… · p. 5
• After event selection, a hybrid reconstruction method (16, 17) combines maximum-likelihood estimation with deep learning, maximizing the likelihood of the observed light pattern under an event hypothesis to determine event properties.Source: Application of deep learning to… · p. 6
• The network approximates Monte Carlo light predictions while encoding detector knowledge and interaction symmetries, intended to use available information more fully despite detector-convolved data.Source: Application of deep learning to… · p. 6
• The selection retains more than 20 times as many events as the previous cascade-based Galactic plane analysis (12), with angular resolution improved by up to a factor of 2 at TeV energies (16).Source: Application of deep learning to… · p. 6
• The ten-year sample contains 59,592 events from 500 GeV to multiple PeV, compared with 1,980 events from seven years in the previous selection (12).Source: Application of deep learning to… · pp. 6–7
• The sample is estimated to contain about 6% atmospheric muons, about 7% astrophysical neutrinos, and 87% atmospheric neutrinos; these fractions are not used directly.Source: Application of deep learning to… · p. 7

Searches for Galactic neutrino emission

• Backgrounds were estimated by scrambling event right ascension, preserving detector-specific artifacts in the data.The techniques were decided a priori and blind to reconstructed event directions; signal-model mismatch could reduce sensitivity but not change p-values.Source: Searches for Galactic neutrino emission · p. 7
• Predefined tests covered diffuse Galactic-plane emission, known Galactic TeV-source catalogs, the Fermi Bubbles, all-sky point-like sources, and a catalog of GeV gamma-ray emitters.Source: Searches for Galactic neutrino emission · pp. 7–8

Galactic plane neutrino searches

• The three diffuse-emission models were converted to spatial templates convolved with detector acceptance and event angular uncertainty, with signal-event count free and spectrum fixed.The analyses made no assumption about absolute model normalization.Source: Galactic plane neutrino searches · pp. 8–9
• 4.48σ after a conservative three-template trial correction, the Galactic-plane searches reject the background-only hypothesis.Before correction, the π0, KRA5, and KRAγ models had significances of 4.71σ, 4.37σ, and 3.96σ, respectively.Source: Galactic plane neutrino searches · p. 9
• After accounting for trials, no individual location in the all-sky point-source search was statistically significant.Some excesses coincided spatially with gamma-ray emitters, including the Crab Nebula, 3C 454.3, and the Cygnus X region.Source: Galactic plane neutrino searches · p. 9

Searches using catalogs of Galactic sources

• Each Galactic-source catalog search reached at least 3σ against background-only, but the excesses were not interpreted as detections because the catalogs overlap high-flux Galactic-plane regions.The three catalogs comprised supernova remnants, pulsar wind nebulae, and other unidentified Galactic sources.Source: Searches using catalogs of Galactic… · p. 10

Implications of Galactic neutrinos

• The KRAγ best-fitting flux normalizations are lower than predicted, potentially indicating a spectral cutoff inconsistent with the assumed 5 PeV and 50 PeV values.Model-to-model flux comparisons depend on the sky regions considered.Source: Implications of Galactic neutrinos · p. 11
• Injecting the best-fitting π0 signal yields median significances consistent with other tested hypotheses within expected statistical fluctuations, including the 3σ excess in Galactic source-catalog searches.The tests simulate a best-fitting signal from one search and examine expected results in other analyses.Source: Implications of Galactic neutrinos · p. 11
• Injecting any tested Galactic source catalog at its observed flux does not recover the π0 or KRAγ results, but angular resolution and catalog size do not constrain broad source populations.Many independently contributing Galactic-plane sources could produce results comparable to diffuse interstellar-medium emission.Source: Implications of Galactic neutrinos · p. 11
• The tests favor diffuse Galactic-plane neutrino emission, but lack the statistical power to distinguish emission models or identify embedded point sources.Many independently contributing Galactic sources could produce a result comparable to diffuse emission.Source: Implications of Galactic neutrinos · p. 11
• At 30 TeV, the Galactic template models contribute ∼6–13% of the previously observed all-sky astrophysical flux.Comparisons are complicated by different spectral assumptions and energy ranges; the all-sky flux is integrated over the sky, while central-plane local contributions are higher.Source: Implications of Galactic neutrinos · pp. 11–12
• The observed excess provides strong evidence that the Milky Way is a high-energy neutrino source, complementing IceCube’s diffuse extragalactic-flux measurement.This supports the understanding of Galactic cosmic-ray interactions established by gamma-ray measurements.Source: Implications of Galactic neutrinos · p. 12

• The supplementary materials accompany “Observation of high-energy neutrinos from the Galactic Plane.”Source: List of Supplementary Materials · p. 28

10 Institut für Physik, Humboldt-Universität zu Berlin, D-12489 Berlin, Germany

• Affiliation 11 is the Faculty of Physics and Astronomy at Ruhr University Bochum in Germany.Source: §10 · p. 31
• Affiliation 12 is the Science Faculty at Université Libre de Bruxelles in Belgium.Source: §10 · p. 31
• Affiliation 13 is the Elementary Particles service at Vrije Universiteit Brussel in Belgium.Source: §10 · p. 31
• Affiliation 14 is the Department of Physics and Laboratory for Particle Physics and Cosmology at Harvard University in Cambridge, Massachusetts.Source: §10 · p. 31
• Affiliation 15 is the Department of Physics at the Massachusetts Institute of Technology in Cambridge, Massachusetts.Source: §10 · p. 31
• Affiliation 16 is the Department of Physics and International Center for Hadron Astrophysics at Chiba University in Japan.Source: §10 · p. 31

27 Physik-department, Technische Universität München, D-85748 Garching, Germany

• The affiliation fragment names the Département de physique nucléaire et corpusculaire, Université de Genève.The supplied text ends with the address prefix CH-1211 Genève.Source: §27 · p. 32

30 Department of Physics and Astronomy, University of California, Irvine, CA 92697, USA

• Affiliations 31–34 name Karlsruhe Institute of Technology, Queen’s University, and the University of Kansas.Source: §30 · p. 32
• Affiliations 35–38 include the University of California, Los Angeles, Université catholique de Louvain, Mercer University, and the University of Wisconsin–Madison.Source: §30 · p. 32
• Affiliations 39–42 list the University of Wisconsin–Madison, the University of Mainz, Marquette University, and the University of Münster.Source: §30 · p. 32
• Affiliations 43–46 identify the University of Delaware, Yale University, the University of Oxford, and Drexel University.Source: §30 · pp. 32–33
• Affiliations 47–50 include South Dakota School of Mines and Technology, the University of Wisconsin–River Falls, the University of Rochester, and the University of Utah.Source: §30 · p. 33
• Affiliations 51–54 name Stockholm University, Stony Brook University, Sungkyunkwan University, and the Institute of Basic Science.Source: §30 · p. 33
• Affiliations 55–58 list Academia Sinica, the University of Alabama, and Pennsylvania State University.Source: §30 · p. 33
• Affiliations 59–64 include Uppsala University, the University of Wuppertal, DESY, the University of Padova, the University of Tokyo, and TU Dortmund University.Source: §30 · pp. 33–34

Event Selection and Reconstruction

• The early CNN selection reduces atmospheric background by about 99.92% relative to the online cascade filter (18,22), while retaining more than half of signal events above 500 GeV.Source: Event Selection and Reconstruction · p. 35
• Final gradient-boosted decision trees reduce atmospheric muon background by eight orders of magnitude, leaving neutrinos dominant in the selected sample.Source: Event Selection and Reconstruction · p. 35
• The cascade selection retains more than 20 times as many events as the previous IceCube analysis (12), with a threshold of about 500 GeV versus several TeV.Source: Event Selection and Reconstruction · p. 35
• Deep-learning tools retain challenging partially contained and uncontained events while maintaining reconstruction quality and reducing background contamination; the sample's energy-dependent angular resolution improves over the previous selection.Source: Event Selection and Reconstruction · pp. 36, 39

Combining maximum-likelihood with deep learning

• The hybrid likelihood reconstruction uses neural networks to approximate pulse-arrival-time distributions at 5160 DOMs, modeling high-dimensional dependencies without prior dimensionality reductions.Source: Combining maximum-likelihood with deep… · p. 39
• The uncertainty estimator performs better than evaluating the likelihood Hessian, but its circularized uncertainties simplify analysis despite asymmetric cascade contours and computational limits on likelihood scans.Source: Combining maximum-likelihood with deep… · p. 41
• Detector systematics degrade angular resolution by about 5% at 1 TeV and 30% at 1 PeV; uncertainty estimates over-cover by about 15% in baseline simulation, with less over-coverage when systematics are varied.Source: Combining maximum-likelihood with deep… · pp. 41–42
• Computed p-values are insensitive to uncertainty-estimate accuracy and independent of mis-modeling from the data-driven search method.Source: Combining maximum-likelihood with deep… · p. 42

Spatial distribution of Galactic Models

• Three Galactic diffuse-neutrino templates draw on gamma-ray spatial and spectral information, with KRAγ models incorporating radially dependent cosmic-ray diffusion.The templates are binned over the sky and analysis energy range for use as spatial and spectral templates.Source: Spatial distribution of Galactic Models · pp. 42–43
• Cascade directional resolution washes out small-scale template structure, making the analysis robust to spatial-signal mis-modeling of up to a few degrees.The smearing can aid detection of poorly characterized extended emission but complicates further dissection of the signal.Source: Spatial distribution of Galactic Models · p. 43
• The observed flux, interpreted as diffuse Galactic emission, is consistent with the TeV–PeV level inferred by gamma-ray observatories, but unresolved-source contributions cannot be constrained.The regional flux comparison is a presentation of sky-integrated results, not independent measurements.Source: Spatial distribution of Galactic Models · p. 44, Figure S8 · p. 45

Maximum Likelihood-Based Searches

• The search uses a data-driven background PDF and right-ascension scrambling to preserve background properties while reducing false discovery of anisotropic sky structure.Scrambling removes signal anisotropy while retaining the influence of systematic uncertainties and background-event statistics.Source: Maximum Likelihood-Based Searches · p. 49

Signal Hypotheses

• Signal hypotheses are separated into spatial and energy components, denoted SSpace and SEnergy.Source: Signal Hypotheses · p. 50
• Signal likelihoods combine energy-spectrum weights with spatial weights based on reconstructed direction, angular uncertainty, and the point-source or template hypothesis.Point-source spatial weights use a von Mises-Fisher distribution; template weights use acceptance-convolved maps smeared by each event’s angular uncertainty.Source: Signal Hypotheses · p. 50

Effective Area

• Effective areas are simulation-derived and sum all neutrino flavors and types assuming equal flavor ratios; flux-to-rate conversion uses energy- and declination-dependent area and solid angle.The reported areas are averaged over declination regions.Source: Effective Area · pp. 50–51

Systematic Uncertainties and their Impact

• Confidence intervals for Galactic-plane model flux normalizations include statistical uncertainties and known systematics by sampling a systematic-parameter realization in every trial.The intervals are constructed by inversion of the likelihood ratio test (59).Source: Systematic Uncertainties and their… · pp. 51–52
• Systematics affect sensitivity by up to 20% and effective area by about 10%, with high-energy angular resolution most affected, especially by glacial-ice modeling.The analysis accounts for these effects in the p-value calculation.Source: Systematic Uncertainties and their… · p. 52
• The fitted flux normalization is susceptible to systematic uncertainties, but the data-driven search method makes the identification of Galactic-plane neutrino emission robust.Its p-values are derived from scrambled experimental data, reducing the impact of unmodeled systematics.Source: Systematic Uncertainties and their… · pp. 51–52

All-Sky Scan

• In the a priori all-sky scan and source-list search, no individual point-like source was statistically significant after accounting for the corresponding analysis’s trial factor; all-sky excesses were consistent with background fluctuations.The all-sky test used solid-angle bins spaced 0.45° apart and entailed a large (~500) trial factor.Source: All-Sky Scan · p. 53, Figure S11 · p. 55

Fermi Bubbles and Source Stacking

• The Fermi Bubbles search yielded a trial-corrected p-value of 6.48 × 10^-2 (1.52σ), consistent with background across the tested cutoffs.The template comprised two 25° circular lobes tangent at the Galactic center, with equal emission probability across each lobe and an E^-2 spectrum with varied cutoffs.Source: Fermi Bubbles and Source Stacking · p. 54

Source List

• A search of 109 a priori selected sky positions found results consistent with background and improved sensitivity in the Southern Sky, while track-based analyses remained much more sensitive in the Northern Sky.It placed 90% confidence flux upper limits on source flux at Eν = 100 TeV.Source: Source List · p. 56