Probing neutrino emission at GeV energies from astrophysical transient events with the IceCube Neutrino Observatory

The study investigates whether transient astrophysical events, notably Gamma Ray Bursts (GRBs) and compact binary mergers, produce neutrinos in the GeV energy range that are detectable by IceCube. Neutrinos’ weak interaction and neutrality preserve directional information and probe regions opaque to photons, tracing hadronic processes. While solar MeV neutrinos and TeV–PeV astrophysical neutrinos (e.g., from TXS 0506+056 and NGC 1068) have been observed, the sub-10 GeV regime remains relatively unexplored for astrophysical sources. IceCube’s DeepCore enables sensitivity down to 0.5–5 GeV, where atmospheric neutrinos and detector noise dominate. By leveraging transient timing and comparing on-time to off-time windows, the study aims to probe this low-energy domain, using GRB 221009A as a first benchmark and outlining improvements to expand future searches.

• Solar neutrinos (MeV) are well established; TeV–PeV astrophysical neutrinos have been associated with sources such as TXS 0506+056 and NGC 1068.
• GRBs are theorized to produce GeV neutrinos via hadronuclear interactions, including below-photosphere production.
• Super-Kamiokande has set upper limits on GeV astrophysical neutrinos but is limited by detector size; KM3NeT will contribute once completed.
• IceCube, optimized for TeV–PeV, can access GeV energies via DeepCore; previous IceCube transient searches have focused on higher energies, with some first constraints at GeV using the ELOWEN selection.
• Related IceCube efforts include searches for neutrinos from solar flares and compact binary mergers, with catalog-based analyses through O3 and ongoing O4 campaigns.

Detector and selection: IceCube’s DeepCore sub-array is used with the ELOWEN selection to target 0.5–5 GeV neutrinos. The initial trigger rate (~1400 Hz) is dominated by atmospheric muons and detector noise (thermal, radioactive, scintillation). Sequential filtering reduces rate to ~0.02 Hz:

1. Containment in DeepCore and veto of higher-energy filters (~15 Hz).
2. Constraint on number of triggered optical modules (~6 Hz) to reject high-energy and partially contained events and additional noise.
3. NoiseEngine correlated-hit filter (~0.2 Hz).
4. Data-quality cuts on depth of interaction, total charge, and spacetime separation between first hits, yielding ~0.02 Hz. The final sample retains ~40% of simulated neutrinos from 0.5–5 GeV following an E^-2 spectrum (GENIE 2.8.6) but remains dominated by detector noise; atmospheric neutrinos are expected at mHz levels.

Transient search strategy: For transients, a counting analysis compares on-time event counts with off-time background windows of identical duration drawn from periods without known transients (novae, GRBs, GWs, solar flares). Significance is determined by the empirical background distribution from 2600 off-time windows per duration.

GRB 221009A analysis: Two windows were tested: (i) 1000 s centered on T0; (ii) 2200 s starting 200 s before T0. Pre-unblinding checks included: (a) verifying an 8-hour pre-window rate matches the expected ~0.02 Hz without anomalous fluctuations; (b) ensuring no spatially localized detector artifacts in candidate distributions. All checks passed before unblinding.

Binary mergers: Catalog-based searches (up to O3 and ongoing O4) adopt a 1000 s standard window and, for binary neutron star mergers, an additional 3 s precursor window. The same off-time background strategy is used; upper limits are set where no excess is found.

Planned improvements:

• Noise rejection: Reassessment of variables across all filters using t-SNE indicates neutrino-dominated clusters, suggesting further separability. Enhanced NoiseEngine configurations combined with a boosted decision tree (BDT) optimized for signal-to-noise (not accuracy) can impose stricter thresholds; for BDT score >0.95, residual noise rate becomes comparable to neutrinos at that stage.
• Direction reconstruction: A first zenith-direction estimate at GeV energies uses hits on the DeepCore string with the most activity, leveraging downward-facing DOM geometry and timing between the first HLC and neighboring DOMs. Two BDTs trained to classify upgoing and downgoing events achieve a balanced accuracy of ~77%; ~40% of events remain unclassified (e.g., horizontal tracks or too few hits). More advanced algorithms are under development.

• GRB 221009A on-time counts are consistent with background for both windows: p = 0.79 (1000 s) and p = 0.81 (2200 s).
• 90% C.L. upper limits on the time-integrated all-flavor neutrino fluence (assuming E^-2 from 0.5–5 GeV; reference energy 1 GeV): 5.3 × 10^3 GeV cm^-2 (1000 s) and 7.9 × 10^3 GeV cm^-2 (2200 s).
• ELOWEN selection reduces trigger rate from ~1400 Hz to ~0.02 Hz while retaining ~40% of targeted low-energy neutrinos, but the final sample remains noise-dominated relative to atmospheric neutrino expectations (mHz level).
• Machine-learning enhancements show promise: a BDT tuned for S/N with threshold >0.95 reduces noise to a level similar to neutrinos at that filter stage; t-SNE reveals regions of strong separation between noise and simulated neutrinos.
• Initial zenith-direction reconstruction at GeV energies yields ~77% balanced accuracy for up/down classification with ~40% unclassified events.

The non-detection for GRB 221009A indicates that, within the 0.5–5 GeV range and assumed E^-2 spectrum, any neutrino emission was below IceCube’s current sensitivity, enabling the first constraints on sub-5 GeV fluence from this exceptionally bright GRB. Using transient timing and empirical off-time backgrounds provides a robust method to suppress irreducible atmospheric and detector noise without requiring spatial reconstruction. The derived limits contribute to constraining hadronuclear models predicting low-energy neutrinos from GRB sub-photospheric regions. Planned noise-reduction and nascent zenith-direction reconstruction will directly improve signal-to-noise and enable incorporating directional information in future transient searches, thereby increasing discovery potential and tightening constraints on theoretical models.

IceCube, via DeepCore and the ELOWEN selection, can probe astrophysical neutrinos in the 0.5–5 GeV range. The first search targeting GRB 221009A found no excess, setting 90% C.L. fluence upper limits and establishing a baseline for GeV-range transient analyses. Ongoing developments—machine-learning-driven noise rejection and initial zenith-direction reconstruction—are expected to significantly enhance sensitivity. With these improvements, IceCube will expand searches to a broader set of transients (including compact binary mergers), improving our understanding of hadronic processes and particle acceleration in astrophysical sources.

• The final ELOWEN sample remains dominated by detector noise at ~0.02 Hz, exceeding expected atmospheric neutrino rates (mHz), limiting sensitivity.
• Lack of full direction reconstruction at GeV energies constrains analyses to time-window counting; current zenith-only classification leaves ~40% of events unclassified and provides no azimuthal information.
• Event topologies at GeV energies are small relative to DOM spacing, reducing reconstruction power.
• Results for GRB 221009A assume a specific spectral shape (E^-2, 0.5–5 GeV), and constraints may vary under different spectral models.
• Binary merger results referenced are upper limits with ongoing analyses; not all source classes or time structures are fully explored yet.