IceCube-Gen2: Next-Gen Neutrino Observatory

• IceCube-Gen2 is a next-generation neutrino observatory that uses a triple-array system (optical, surface, and radio) to significantly boost neutrino and cosmic-ray detection rates.
• Its optical array features multi-PMT modules and a shift to timestamp-first data flow, improving detection efficiency and directional resolution.
• The combined observatory design enables an order-of-magnitude increase in cosmic-neutrino rates and enhanced steady-source sensitivity with cross-calibrated multi-messenger data.

IceCube-Gen2 is the planned next-generation extension of the IceCube Neutrino Observatory at the South Pole. In its mature design literature, it is a co-located three-component facility consisting of an enlarged in-ice optical Cherenkov array for TeV–PeV neutrinos, a surface air-shower array for vetoing and cosmic-ray measurements, and a sparse in-ice radio array for ultra-high-energy neutrinos extending into the EeV regime. Across successive design papers, the observatory is described as increasing the observed cosmic-neutrino rate by about an order of magnitude, improving steady-source sensitivity by about a factor of five, and combining neutrino astronomy with a hybrid cosmic-ray program based on simultaneous measurements of electromagnetic shower components, low-energy surface muons, and high-energy muons in deep ice (Clark, 2021, Collaboration et al., 2020, Ishihara, 2023).

1. Observatory concept and architectural evolution

The published concept of IceCube-Gen2 has evolved from an early 10 km$^3$ vision to a later baseline centered on about 8 km$^3$ of instrumented ice, while preserving the same fundamental architecture: a high-energy optical array, a surface array, and a radio array (Collaboration et al., 2014, Clark, 2021). Later overview papers describe the optical array as using about 10,000 optical sensors and the surface and radio systems as co-deployed over the same South Pole site, with the radio array extending over more than 400 km$^2$ or roughly 500 km$^2$, depending on the design stage being discussed (Ishihara, 2023, Glaser, 10 Jul 2025).

For the optical component, a recurrent reference configuration is a 120-string extension around IceCube in a “Sunflower” layout with about 240 m lateral spacing. Geometry optimization studies varied the inter-string spacing from 150 m to 350 m and found that 200–280 m gives the best overall point-source discovery potential for an $E^{-2}$ spectrum, with 240 m repeatedly used as an illustrative reference. The Sunflower layout is generated from the polar-spiral prescription

$$r = s \sqrt{n}, \qquad \phi = \frac{2\pi}{g^2}n, \qquad g=\frac{1+\sqrt{5}}{2},$$

with deployment constraints excluding parts of the South Pole site (Omeliukh, 2021).

The facility is also conceived as a phased build. One deployment study models the in-ice optical array over approximately seven Antarctic summer seasons and emphasizes that substantial gains appear before full completion. At the halfway point of that deployment, the simulated optical array already reaches approximately twice the effective area of IceCube for horizontally throughgoing events and about 50% better angular resolution (Clark et al., 2021). This suggests that IceCube-Gen2 is designed not only as a final detector geometry but also as an observatory whose scientific reach grows continuously during construction.

2. In-ice optical array and optical-module technology

The optical array is the TeV–PeV core of IceCube-Gen2. One baseline description gives 120 new strings, each carrying about 80 optical modules between depths of 1340 m and 2600 m with 17 m vertical spacing, instrumenting about 8 km$^3$ of glacial ice (Clark, 2021). A geometry-optimization study uses a closely related configuration with modules between 1325 m and 2575 m at 16 m spacing, showing that the exact vertical prescription depends on the analysis context (Omeliukh, 2021). The recurring design objective is to exploit the long optical absorption lengths of South Pole ice to enlarge the detector footprint without losing astrophysical sensitivity at high energies.

Sensor development departs from IceCube’s single large PMT DOM architecture. Proceedings on the next-generation telescope describe Gen2 optical sensors as elongated multi-PMT modules using 4-inch PMTs and improved electronics, with combined photon detection efficiency improved by close to a factor of four (Ishihara, 2023). Electronics studies for a prototype optical module describe 16- or 18-PMT concepts, each PMT being digitized by a 2-channel, 12-bit ADC at 60 MSps, with low- and high-gain paths to cover a dynamic range from roughly 1 to 10,000 photoelectrons. The same prototype architecture includes LED flashers, an FPGA for in-module local coincidence, and onboard $\mu$SD flash memory for hit buffering before data are requested from the surface (Griffin, 2023).

A major architectural change is the shift from continuous waveform transmission to timestamp-first data flow. In the prototype design, only timestamp information is continuously transmitted to the surface, while full hit data are buffered in-ice and retrieved on demand when an array-level trigger fires. The paper estimates an average compressed data rate of about 500 kbps per Gen2 DOM and adopts a 32 GB hitspool depth to buffer approximately one week of data, specifically to enable sub-threshold supernova searches (Griffin, 2023). This DAQ model is central to scaling the observatory to approximately 10,000 optical sensors.

Multi-PMT optical modules also alter the detector-response model. A machine-learning study adopts a 16-PMT baseline prototype with sixteen 4-inch PMTs arranged symmetrically into upper and lower hemispheres, each hemisphere containing four polar and four equatorial PMTs. Gel-pad optics around the photocathodes enhance detection via total internal reflection, and these multi-PMT modules yield up to a factor of four improvement in average effective area over current IceCube DOMs (Carbonell et al., 10 Jul 2025). That increased angular coverage is one reason later Gen2 papers emphasize improved cascade reconstruction and more complete directional information at the module level.

3. Surface array, air-shower measurements, and atmospheric-veto function

The surface array is simultaneously a veto system for neutrino astronomy and a cosmic-ray detector. Across design papers, each surface station is consistently described as comprising eight scintillation panels and three radio antennas, co-located above optical strings and digitized through a field-hub architecture shared with the in-ice system (Coleman, 2023, Schroeder, 27 Feb 2025). Some proceedings describe the scintillators as arranged as four pairs with 5 m panel separation, while others describe a triangular geometry grouped in pairs along three arms; these differences indicate evolving engineering descriptions of the same hybrid-station concept (Coleman et al., 2021, Coleman, 2023).

Published surface-array footprints vary between about 6 km$^2$ and approximately 8 km$^2$, and station counts are described as approximately 160 or 162 depending on whether the enhancement of the existing IceTop region is included in the quoted baseline (Coleman, 2023) (Clark et al., 2021)? no surface counts there. A later technical summary ties the surface array directly to the 120 new optical strings and states that, with the IceTop enhancement, the full surface footprint will host approximately 160 stations (Schröder, 8 Jul 2025). This suggests that station count and area are design-level quantities rather than immutable final values.

The scintillation subsystem sets the low-energy threshold and the veto performance. Several studies converge on a near-vertical full-efficiency threshold around 0.5 PeV. One simulation paper defines a detected event as at least five scintillator panels across the array registering at least 0.5 MIP and reports full efficiency at $^3$0 PeV for near-vertical showers (Coleman et al., 2021). A later design note states that the station layout provides sensitivity to $^3$1 PeV air showers with a simple 3-panel trigger, while another 2025 proceedings paper again quotes full efficiency around 0.5 PeV for vertical and mildly inclined protons and photons under the five-panel, half-MIP condition (Coleman, 2023, Schröder, 8 Jul 2025). The common point is that the scintillator layer is intended to preserve a sub-PeV veto threshold even as the detector footprint expands.

The radio component of the surface array provides calorimetric information on the electromagnetic cascade and sensitivity to the depth of shower maximum, $^3$2. Surface-array papers consistently use SKALA v2 antennas operating in the 70–350 MHz band, mounted above the snow together with the scintillators to avoid the threshold drift that affected IceTop (Schroeder, 2021, Schroeder, 27 Feb 2025). The hybrid surface design is therefore explicitly tied to several science goals: spectrum and anisotropy measurements across the PeV–EeV regime, composition reconstruction through the combination of electromagnetic and muonic observables, and hadronic-interaction studies including the “muon puzzle” (Coleman et al., 2021, Schröder, 8 Jul 2025).

The key gain relative to IceTop is the enlarged surface–deep coincidence aperture. Multiple papers state that the aperture for coincident surface–in-ice events grows by about a factor of 30, driven by the larger area and wider zenith-angle reach (Coleman, 2023, Schroeder, 27 Feb 2025). One proceedings paper further quantifies the neutrino-astronomy consequence: in a ten-year exposure, the surface veto increases the total number of astrophysical neutrinos identified in the southern sky by $^3$3 events, and by about 100 TeV deposited energy the southern-sky diffuse analysis becomes effectively background-free (Coleman, 2023). A common misconception is therefore that the surface array is a secondary add-on for cosmic rays alone; in the Gen2 design literature, it is also a primary background-rejection instrument for the optical neutrino program.

4. In-ice radio array and the ultra-high-energy extension

The radio array is the EeV-scale component of IceCube-Gen2. Its function is to detect Askaryan radio emission from neutrino-induced showers in polar ice, thereby extending the observatory’s sensitivity to energies where optical Cherenkov instrumentation becomes cost-inefficient (Hallmann et al., 2021, Glaser, 10 Jul 2025). Early sensitivity studies examine an array of 313 stations over about 500 km$^3$4, whereas a more recent design paper describes 361 stations on a square grid with 1.24 km spacing, alternating between shallow-only and hybrid stations (Hallmann et al., 2021, Glaser, 10 Jul 2025). This is one of the clearest examples of evolving baseline specifications across the literature.

In the newer design, 197 shallow-only stations each carry 8 channels near the surface, while 164 hybrid stations each carry 24 channels and add a phased array at 150 m depth, additional antennas on the main string, and two supplementary strings (Glaser, 10 Jul 2025). Earlier station-level studies give more explicit channel composition: shallow stations with four downward-facing LPDAs, three upward-facing LPDAs for cosmic-ray tagging, and one vertically polarized dipole at 15 m depth, plus a deep component using a four-antenna phased array distributed along a 200 m string (Hallmann et al., 2021). Both descriptions share the same logic: shallow antennas provide wide acceptance, polarization information, and air-shower rejection, while deeper antennas mitigate firn-refraction effects and lower the trigger threshold.

The radio array’s performance is controlled by neutrino interaction lengths, Earth absorption, and the propagation of impulsive Askaryan signals in firn and deep ice. A recent design paper writes the interaction length as

$^3$5

and the Earth-transmission factor as

$^3$6

with $^3$7 the chord depth through Earth (Glaser, 10 Jul 2025). The same paper emphasizes that the radio array is designed for discovery and characterization above approximately 10 PeV and that it is expected to increase sensitivity beyond 100 PeV by at least an order of magnitude over existing arrays (Glaser, 10 Jul 2025).

Performance projections in the earlier full-array radio sensitivity study are more explicit about event yields. For an unbroken astrophysical spectrum following $^3$8, the radio detector is expected to measure 74 neutrinos in ten years, concentrated between $^3$9 and $^2$0 eV. For cosmogenic fluxes, the same study gives about 240 detected neutrinos in ten years for a Telescope Array best-fit UHECR model and about 20 for an Auger-fit model with 10% proton fraction (Hallmann et al., 2021). More recent forecast work adds that the zenith-dependent absorption signal should permit a $^2$1 cross-section measurement above 100 PeV to approximately 50% precision in 10 years, and that combined flavor-sensitive analyses can distinguish pion-decay, muon-damped, and neutron-decay scenarios at $^2$2 confidence level in a high-flux case with about 180 events in 10 years (Glaser, 10 Jul 2025).

A second misconception concerns background levels. The radio literature in the design set repeatedly stresses that the array is intended to operate near-background-free at ultra-high energies, but it does not treat background rejection as trivial. Upward-facing LPDAs, multi-station coincidence, hybrid station cross-validation, and careful modeling of thermal noise and air-shower radio emission are all explicit parts of the design, and several DAQ and trigger parameters are still being optimized (Hallmann et al., 2021, Glaser, 10 Jul 2025).

5. Simulation, reconstruction, and machine-learning infrastructure

IceCube-Gen2 is unusual among large observatories in how centrally machine learning appears in subsystem design rather than only in high-level analysis. A 2025 methods paper focuses on three concrete tasks for the optical array: DOM-level photon-acceptance simulation, probabilistic reconstruction of $^2$3 charged-current tracks, and pulse-level noise cleaning (Carbonell et al., 10 Jul 2025).

For optical-module simulation, the paper replaces the symmetry-limited analytical response $^2$4 used in PPC with a neural-network model trained inside OMNNSim/OMSim on about 20 billion isotropically generated photons with flat wavelength distribution between 270 and 700 nm. The network combines a PMT-relative branch, which exploits polar/equatorial symmetries, with an absolute-input branch, which learns symmetry-breaking effects such as cable shadowing. It reproduces Geant4 angular structures missed by the analytical model and processes one million photons in 0.3 s on a GPU, corresponding to an approximate 200–300$^2$5 speedup relative to Geant4 on CPU (Carbonell et al., 10 Jul 2025).

For event reconstruction, the same paper uses GraphNeT to generate 6.5 million $^2$6 CC events between 1 TeV and 50 PeV and studies a GNN-plus-transformer architecture based on the winning IceCube Kaggle approach. To improve scalability, the standard transformer is replaced with Performer attention, yielding a 4$^2$7 training speedup. Two probabilistic direction estimators are compared: a 3D von Mises–Fisher model and conditional normalizing flows. Above 10 TeV, the flow-based contours match empirical coverage to within 5%; at lower energies the method undercovers, which the authors interpret as overfitting that could be mitigated by low-energy augmentation or weighting. To map the flow posterior to a scalar uncertainty proxy, they use the large-$^2$8 relation

$^2$9

With a quality cut $^2$0, conditional flows give the best directional performance, especially for starting events (Carbonell et al., 10 Jul 2025).

For noise cleaning, the same paper trains a DynEdge graph neural network in GraphNeT to classify individual pulses as physics or noise on point-cloud graphs built from PMT position, PMT orientation, pulse charge, and pulse time. On roughly two million simulated $^2$1 CC and NC events, the GNN suppresses more than 99% of noise pulses up to $^2$2 pulses per event, compared with roughly 70% for the scaled Seeded RT baseline (Carbonell et al., 10 Jul 2025). The paper is explicit that Gen2 waveform simulation for PMTs is not yet available, so mixed physics/noise hits inside single waveforms remain an open validation issue.

Outside the explicitly machine-learning work, Gen2 subsystem papers also show a heterogeneous simulation ecosystem. Optical-array studies use CLSim and likelihood-based track reconstruction such as SplineMPE (Omeliukh, 2021). Surface-array studies use CORSIKA with FLUKA and Sibyll 2.3d, Geant4 for scintillators, CoREAS for radio emission, and PROPOSAL for muon transport (Coleman et al., 2021). Radio-array sensitivity studies use NuRadioMC and NuRadioReco with full Askaryan-signal propagation and station-response modeling (Hallmann et al., 2021). In practice, IceCube-Gen2 is therefore not one simulation chain but a federation of detector-specific frameworks linked by common geometry, calibration, and event-reconstruction goals.

6. Scientific reach and projected performance

The central scientific aim of IceCube-Gen2 is to convert IceCube’s discovery-era neutrino astronomy into a higher-statistics, broader-band program. General observatory papers consistently state two headline gains for the optical array: an order-of-magnitude increase in the rate of observed cosmic neutrinos and a factor-of-five improvement in steady-source sensitivity (Clark, 2021, Collaboration et al., 2020). For high-quality through-going tracks, one paper quotes an effective-area increase of about $^2$3 and a median angular error about $^2$4 smaller than IceCube, while another later overview describes an approximate three-fold improvement in muon-track angular resolution; these are different projected figures under different selections, but both point to materially better localization of astrophysical neutrino sources (Clark, 2021, Ishihara, 2023).

The time-domain science case is similarly prominent. The 2021 observatory overview states that a 156-day TXS 0506+056-like flare would be observed at $^2$5 significance in Gen2, even without a gamma-ray counterpart (Clark, 2021). Another design paper projects that NGC 1068 would reach approximately $^2$6, enabling precise spectral and variability studies (Ishihara, 2023). More generally, the facility is designed to improve real-time multi-messenger follow-up through better localization, higher alert rates, and a broader energy baseline than IceCube alone.

The cosmic-ray science program is equally broad. Surface-array papers describe measurements from several $^2$7 TeV to a few EeV, including energy spectrum, mass composition, large-scale anisotropy, prompt muon production, hadronic-interaction studies beyond LHC phase space, and searches for PeV photons (Coleman, 2023, Schröder, 8 Jul 2025). Composition sensitivity relies on the complementarity of $^2$8, electromagnetic shower energy, surface muons, and high-energy in-ice muons. One design study summarizes the combined mass dependence through the Heitler–Matthews scaling

$^2$9

and reports that the combined observables achieve approximately $E^{-2}$0 separation between proton and iron populations under expected resolutions (Coleman, 2023). A 2025 surface-array proceedings paper further states that, with 10 years of exposure, the array achieves sensitivity to dipole amplitudes at the percent level in the 10’s of PeV range (Schröder, 8 Jul 2025).

At the highest energies, the radio array is intended to determine whether the astrophysical spectrum continues into the EeV regime and whether a cosmogenic component is present. The relevant design papers emphasize discovery and characterization rather than only upper limits: diffuse-flux discovery within a few years for many benchmark models, point-source multiplet searches above 100 PeV, cross-section measurements, and flavor studies when optical and radio data are interpreted together (Glaser, 10 Jul 2025, Hallmann et al., 2021). A plausible implication is that IceCube-Gen2 is being framed not as three isolated instruments but as a genuinely wide-band neutrino observatory whose physics program depends on cross-calibration between its optical, surface, and radio subsystems.

7. Calibration, deployment path, and unresolved technical questions

Because IceCube-Gen2 is sparse and geographically large, calibration of the Antarctic medium is unusually consequential. A 2025 calibration paper argues that the larger footprint increases the importance of “ice tilt,” the undulation of layers of constant optical properties across the detector volume. To address this, the collaboration is developing a co-deployed laser dust logger using a modified 405 nm POCAM emitter and a 16-PMT Low-Noise Optical Module receiver. The system is designed to resolve stratigraphic features with FWHM of about 20 cm, and the prototype beam achieved a measured vertical divergence of $E^{-2}$1, well below the $E^{-2}$2 requirement derived from PPC-based simulations (Eimer et al., 11 Jul 2025). The instrument is scheduled for testing during the IceCube Upgrade deployment in the 2025/26 austral summer, explicitly as a pathfinder for Gen2 (Eimer et al., 11 Jul 2025).

The IceCube Upgrade itself functions as a broader technology demonstrator. Gen2 overview papers describe the Upgrade as a testbed for pixelated optical modules, calibration concepts, and improved ice modeling, including planned reduction of drill-hole bubble columns through hole degassing (Clark, 2021). Prototype surface stations have also operated since 2020, and several years of South Pole operation are cited as evidence that elevated scintillators and antennas successfully avoid local snow accumulation (Schroeder, 2021, Schröder, 8 Jul 2025). For the radio array, RNO-G is explicitly used as a development site for DAQ and algorithm maturation, including embedded neural-network triggering (Glaser, 10 Jul 2025).

Several unresolved issues appear consistently across the literature. Radio-station counts, spacings, and channel layouts have changed between design studies, indicating continuing optimization rather than a frozen deployment blueprint (Hallmann et al., 2021, Glaser, 10 Jul 2025). Surface-array footprint estimates differ between approximately 6 km$E^{-2}$3 and approximately 8 km$E^{-2}$4, again reflecting design evolution rather than contradiction (Coleman, 2023, Schroeder, 27 Feb 2025). In the ML domain, low-energy undercoverage in conditional-flow reconstruction and the absence of Gen2-specific waveform simulation for pulse cleaning are explicitly acknowledged limitations (Carbonell et al., 10 Jul 2025). In optical calibration, the dust-logger paper does not yet specify the full mapping from measured impurity stratigraphy to absorption and scattering parameters, only that the logging signal is proportional to impurity content (Eimer et al., 11 Jul 2025).

These caveats are important because Gen2 performance projections are often quoted from different analysis stages, geometry assumptions, and subsystem baselines. The observatory is therefore best understood as a mature but still evolving technical program: the three-array concept, the South Pole site strategy, and the scientific rationale are stable, whereas several detailed engineering choices remain under refinement in response to calibration, logistics, and simulation results (Collaboration et al., 2020, Glaser, 10 Jul 2025).