ICEMAN: IceCube Multi-Flavor Neutrino Dataset

Updated 6 July 2026

The ICEMAN dataset is a federated, full-sky sample assembled from DNN-based cascades, ESTES, and Northern Tracks to measure Galactic neutrino flux and search for PeVatrons.
Enhanced methodologies such as updated cascade reconstruction, SPICE FTP-v3 ice modeling, and GBDT-based per-event PSFs improve angular and energy resolutions significantly.
A unified likelihood framework incorporating multi-topology data increases source-search sensitivity by over 20%, setting the stage for joint flavor and spectral inferences.

Searching arXiv for recent ICEMAN and related IceCube flavor papers. IceCube Multi-Flavor Astrophysical Neutrino sample (ICEMAN) denotes a 12.3-year, full-sky, all-flavor IceCube dataset assembled from three largely independent neutrino samples with different event morphologies: a DNN-based cascade selection, the Enhanced Starting Track Event Selection (ESTES), and the Northern Track sample. In published IceCube proceedings, ICEMAN is presented as a maximally sensitive, multi-flavor, multi-topology sample for measuring the Galactic plane neutrino flux and for searches for Galactic PeVatrons, including individual-source, stacking, and extended-emission analyses. Its recent formulation also incorporates improved shower reconstruction, updated ice modeling, and machine-learning-based per-event point-spread functions (PSFs) for cascades (Thiesmeyer et al., 11 Jul 2025, Seen et al., 10 Jul 2025).

1. Operational definition and sample construction

ICEMAN is defined operationally as the combination of three component samples that contribute complementary sky coverage, angular resolution, energy resolution, and flavor sensitivity. In this usage, “multi-flavor” refers to the fact that cascades include all flavors via neutral-current interactions plus $\nu_e$ and $\nu_\tau$ charged-current interactions, while tracks are dominated by $\nu_\mu$ charged-current interactions and starting tracks also retain some $\nu_\tau$ content. “Multi-topology” refers to the simultaneous use of cascades, starting tracks, and through-going tracks (Thiesmeyer et al., 11 Jul 2025).

Component	Time span / size	Role and stated performance
DNN Cascades (DNNC)	13 May 2011 to 28 November 2023; 85,199 events	Cascade-like, full-sky; at 10 TeV median angular resolution $\sim 10^\circ$ , median energy resolution $3\%$
ESTES	13 May 2011 to 28 November 2023; 11,755 events	Starting muon tracks, full-sky; at 10 TeV median angular uncertainty about $1^\circ$ , median energy uncertainty about $30\%$
Northern Tracks (NT)	1 June 2010 to 28 November 2023; 13 years	Through-going muon tracks from the Northern sky; at 10 TeV median angular resolution about $0.5^\circ$ , median energy resolution about $80\%$

The three components are not simply concatenated. In the combined sample, overlaps between the different datasets are removed. The largest overlaps are between ESTES and Northern Tracks, and overlapping events between those two selections are kept in the starting-track set to maximize sensitivity. This construction makes ICEMAN a full-sky dataset even though the Northern Track component is restricted to zenith angles above $\nu_\tau$ 0, because the cascade and starting-track components remain full-sky (Thiesmeyer et al., 11 Jul 2025).

A common misconception is to treat ICEMAN as a single contained-vertex sample analogous to HESE or MESE. In the published ICEMAN literature, it is instead a federated sample built from three selections with different morphologies and background-rejection strategies. That distinction is central to its intended use: not precision flavor-ratio inference from one homogeneous event class, but a source-search dataset that exploits complementary channels in a unified likelihood (Seen et al., 10 Jul 2025).

2. Reconstruction, topology handling, and per-event spatial modeling

The most technically distinctive part of the current ICEMAN program is the updated cascade component. The proceedings on improved spatial modeling start from an existing 10-year DNN-based cascade selection and update it by applying the same cuts as the previous selection while removing events with predicted angular error larger than $\nu_\tau$ 1, thereby defining “DNN Cascades ICEMAN.” With updated energy reconstructions, the same 10 years of data yield $\nu_\tau$ 2 more events than the original DNN Cascades sample (Seen et al., 10 Jul 2025).

The cascade reconstruction uses the SPICE Full Tilt Parameterization v3 (SPICE FTP-v3), described there as “our best understanding of the detector ice,” together with two maximum-likelihood reconstructions: Monopod, which assumes a single cascade, and Taupede, which assumes a double-cascade morphology. Taupede uses the Monopod result as a seed, and the better fit is retained as the PreferredFit. The stated consequence is that, by using the SPICE FTP-v3 ice model and the PreferredFit reconstruction, biases in azimuth decrease (Seen et al., 10 Jul 2025).

Per-event spatial uncertainty is modeled with a mixed PSF parameterization. For small angular uncertainties, $\nu_\tau$ 3, the spatial term uses a 2D Gaussian on the sphere; for larger uncertainties, $\nu_\tau$ 4, it uses a von Mises–Fisher distribution in order to characterize the tail of the distribution more accurately. The per-event PSF width is predicted by a gradient-boosted decision tree trained with a loss function derived from the negative log-likelihood of the vMF form and reparameterized through $\nu_\tau$ 5 and $\nu_\tau$ 6 (Seen et al., 10 Jul 2025).

The GBDT uses 14 reconstructed quantities from the PreferredFit, including reconstructed vertex coordinates, zenith, azimuth, energy, total deposited charge, several reconstruction-quality metrics, the number of degrees of freedom, distances to the closest string and detector edge, and the length between the two cascades. The training sample comprises 5,718,881 Monte Carlo events, subdivided into 2,570,907 $\nu_\tau$ 7, 1,737,306 $\nu_\tau$ 8, and 1,410,668 $\nu_\tau$ 9 events, simulated across three energy ranges from 100 GeV to 100 PeV and reweighted proportional to an $\nu_\mu$ 0 spectrum. The proceedings note early stopping after 548 iterations and identify reconstructed depth $\nu_\mu$ 1 as the most important feature, consistent with the strong depth dependence of ice properties (Seen et al., 10 Jul 2025).

Several reconstruction-performance numbers are explicitly quoted. Using Monte Carlo with the GBDT-reconstructed PSF, the best median PSF is $\nu_\mu$ 2 degrees in the 290 TeV to 13.8 PeV energy range. For the final DNN Cascades ICEMAN sample after cuts, the smallest median angular resolution is six degrees. The proceedings further report point-source sensitivity and discovery potential for DNN Cascades ICEMAN alone on the order of $\nu_\mu$ 3, assuming an $\nu_\mu$ 4 single-power-law spectrum at 100 TeV (Seen et al., 10 Jul 2025). This suggests that improved reconstruction, not only increased exposure, is a primary driver of the cascade component’s gain in source-search utility.

3. Likelihood architecture and Galactic-plane flux measurement

The first fully articulated physics use of ICEMAN in the current literature is an updated Galactic-plane neutrino analysis. That work adopts an unbinned maximum-likelihood method using the reconstructed direction, energy, and angular uncertainty of each neutrino candidate. The signal model is template-based: Galactic emission templates are acceptance-weighted with the effective area of each ICEMAN sub-sample and then smeared according to the relevant angular resolution. For the examples shown in the proceedings, the smearing scale is $\nu_\mu$ 5 for track samples and $\nu_\mu$ 6 for DNN cascades (Thiesmeyer et al., 11 Jul 2025).

The analysis considers four Galactic templates: Fermi $\nu_\mu$ 7, KRA $\nu_\mu$ 8, KRA $\nu_\mu$ 9, and CRINGE. For the Fermi $\nu_\tau$ 0 template, the energy PDF is constructed by weighting Monte Carlo with a single power law of spectral index $\nu_\tau$ 1; for KRA $\nu_\tau$ 2, KRA $\nu_\tau$ 3, and CRINGE, the corresponding sky-averaged predicted neutrino spectra are used. Background is treated empirically as a declination- and energy-dependent mixture of atmospheric muons, atmospheric neutrinos, and diffuse isotropic astrophysical neutrinos, with a signal-subtraction construction to avoid bias from Galactic emission already present in the data (Thiesmeyer et al., 11 Jul 2025).

The proceedings write the background–signal decomposition as

$\nu_\tau$ 4

and define the unbinned likelihood as

$\nu_\tau$ 5

The test statistic is then

$\nu_\tau$ 6

This is a standard IceCube likelihood architecture, but here it is instantiated with the combined ICEMAN event set rather than a single topology (Thiesmeyer et al., 11 Jul 2025).

The published sensitivities and discovery potentials are:

Template	Sensitivity	$\nu_\tau$ 7 discovery potential
Fermi $\nu_\tau$ 8	$\nu_\tau$ 9 per flavor at 100 TeV	$\sim 10^\circ$ 0 per flavor at 100 TeV
KRA $\sim 10^\circ$ 1	$\sim 10^\circ$ 2 model flux units	$\sim 10^\circ$ 3 model flux units
KRA $\sim 10^\circ$ 4	$\sim 10^\circ$ 5 model flux units	$\sim 10^\circ$ 6 model flux units
CRINGE	$\sim 10^\circ$ 7 model flux units	$\sim 10^\circ$ 8 model flux units

Relative to the previous Galactic-plane analysis, the proceedings state that the sensitivities have improved by over 20%. When the best-fit Fermi $\sim 10^\circ$ 9 flux from the earlier Science Galactic-plane result is injected into pseudo-experiments, the median local significance expected with ICEMAN is $3\%$ 0 for the Fermi $3\%$ 1 template, $3\%$ 2 for KRA $3\%$ 3, $3\%$ 4 for KRA $3\%$ 5, and $3\%$ 6 for CRINGE (Thiesmeyer et al., 11 Jul 2025).

4. Galactic PeVatron searches, source classes, and extended emission

ICEMAN is explicitly framed as a Galactic PeVatron search sample. The source-search program discussed in the proceedings has two principal modes: an individual-source search for neutrinos coincident with $3\%$ 7 TeV $3\%$ 8-ray emitters, and a stacking analysis that groups candidates into five classes: supernova remnants, pulsar wind nebulae, microquasars, unidentified sources, and pulsars with nearby molecular clouds or supernova remnants (Seen et al., 10 Jul 2025).

This program is motivated by the recent detection of multi-PeV photons from the Cygnus region and by catalogues of Galactic $3\%$ 9-ray sources above 100 TeV from LHAASO, HAWC, and HESS. Within the ICEMAN proceedings, the Cygnus Cocoon is the most developed extended-source case study. A conservative neutrino flux is derived assuming $1^\circ$ 0 $1^\circ$ 1-interaction from the LHAASO spectrum and adopting a $1^\circ$ 2 bubble morphology consistent with the LHAASO report, while HAWC is quoted as reporting a $1^\circ$ 3 extension (Seen et al., 10 Jul 2025).

The three ICEMAN components contribute differently to such a search. For ESTES, the proceedings state that no Cygnus source neutrino events above the atmospheric background are expected in 12 years of data. For DNN Cascades, a few events greater than 10 TeV are expected where the Cygnus source contribution dominates the background emission. For Northern Tracks, the Cygnus source contribution begins to dominate the background at energies above 30 TeV, and the Northern Tracks dataset provides most of the point-source sensitivity in the northern sky. Because cascade angular resolution is comparable to the $1^\circ$ 4– $1^\circ$ 5 source size, localizing events to a specific sub-source is challenging, but the cascade component can still improve sensitivity to extended emission (Seen et al., 10 Jul 2025).

This division of labor clarifies the physics logic of ICEMAN. Northern through-going tracks dominate sharp point-source localization in the northern hemisphere; starting tracks extend track sensitivity to the full sky with a smaller effective area but higher containment purity; cascades provide all-flavor reach, southern-sky coverage, and particular leverage for extended Galactic structures. The “multi-flavor” label is therefore inseparable from a “multi-analysis” strategy: each topology class supplies a different projection of the same astrophysical sky (Seen et al., 10 Jul 2025).

5. Relation to IceCube’s broader multi-flavor astrophysics program

ICEMAN emerged within a broader IceCube effort to build contained and all-flavor samples that can jointly constrain diffuse spectra, flavor composition, and source populations. In that wider program, the MESE contained-event analysis based on 11.4 years of data reports best-fit astrophysical flavor fractions at Earth of

$1^\circ$ 6

and states that, for the first time, IceCube can exclude a zero fraction for each flavor at more than 68% confidence; zero $1^\circ$ 7 is rejected at 98.7% CL and zero $1^\circ$ 8 at 91.9% CL (V. et al., 9 Jul 2025). This is not an ICEMAN result, but it is methodologically adjacent: it uses contained vertices, all three flavors, and explicit morphology classes.

Likewise, the MESE diffuse-spectrum measurement from 1 TeV to 10 PeV reports strong evidence for structure in the spectrum beyond a single power law with a significance of $1^\circ$ 9. Its preferred broken-power-law fit has

$30\%$ 0

corresponding to $30\%$ 1 TeV (Basu et al., 8 Jul 2025). This result matters for ICEMAN because ICEMAN source searches use energy PDFs, and the broader IceCube spectral program now disfavors a universal unbroken diffuse power law across the full TeV–PeV interval.

The HESE program provides a parallel high-energy reference. The 12-year HESE flavor analysis, based on 97 events above 60 TeV, finds a best-fit Earthly flavor composition

$30\%$ 2

again consistent with a near-democratic mixture and explicitly incorporating double-cascade $30\%$ 3 information (Lad et al., 9 Jul 2025). The latest IceCube Observatory summary then places these efforts in one framework by identifying, for multi-flavor astrophysical inference, the following ingredients: a contained, all-flavor starting-event selection; a combined topology analysis with tracks, cascades, and double cascades; a multi-component likelihood fit including astrophysical plus conventional and prompt atmospheric fluxes; and a systematics framework that will be improved by the IceCube Upgrade and Gen2 (Eeden, 20 Apr 2026).

Taken together, these analyses situate ICEMAN as the source-search arm of a larger multi-flavor architecture. A plausible implication is that future ICEMAN analyses will increasingly share spectral models, nuisance-parameter treatments, and topology-response calibrations with MESE and HESE-style contained-sample fits.

6. Systematics, detector modeling, and future extensions

The ICEMAN literature emphasizes systematics reduction more than a full published systematic budget. Its central methodological claim is that accurate modeling of the detector ice and detector calibrations leads to improved shower reconstruction and improved angular resolutions. In the cascade proceedings this is expressed through the use of SPICE FTP-v3, PreferredFit reconstruction, and GBDT-based per-event PSFs, with the explicit statement that azimuth biases decrease under the updated model (Seen et al., 10 Jul 2025).

The general detector context is equally relevant. IceCube detects GeV-to-PeV+ neutrinos in about $30\%$ 4 of glacial ice instrumented with 5160 Digital Optical Modules on 86 strings, and the detector’s scientific reach has been tied to increasingly detailed characterization of the glacial ice together with fast approximations of Cherenkov-yield expectations. The IceCube Upgrade will add seven additional strings in a dense infill configuration, multi-PMT optical modules, improved calibration devices, and new sensor prototypes; its denser spacing is intended to extend sensitivity to lower neutrino energies and further constrain neutrino oscillation parameters. IceCube Gen2 is described as increasing the effective volume by nearly an order of magnitude (Yuan, 2022).

In the observatory outlook most directly relevant to multi-flavor astrophysics, the Upgrade is expected to enhance sensitivity to lower-energy neutrinos and reduce systematic uncertainties, while IceCube-Gen2 is expected to expand the detector volume, increase the neutrino detection rate, and extend the energy reach. In the more detailed summary accompanying those observatory results, Gen2 is further described as expanding the instrumented volume by a factor of $30\%$ 5, increasing the annual detection rate of cosmic neutrinos by roughly an order of magnitude, detecting sources up to five times fainter, and extending the energy reach to EeV (Eeden, 20 Apr 2026).

For ICEMAN specifically, the most immediate future direction is combinatorial rather than purely instrumental. Published proceedings already point toward combining the HESE sample with the northern track and cascade samples for a joint flavor measurement, and toward using optimized summary statistics or machine-learning-based particle identification to overcome sparse Monte Carlo statistics in $30\%$ 6-sensitive channels (Lad et al., 9 Jul 2025). That trajectory suggests an eventual convergence between today’s source-optimized ICEMAN construction and the broader IceCube program of unified multi-flavor, multi-topology likelihood analyses.