WavePID: Low-energy flavor identification using single-PMT time series in IceCube

Technical Summary: WavePID: Low-energy flavor identification using single-PMT time series in IceCube

Problem Statement
The IceCube Neutrino Observatory relies on event morphology to distinguish between neutrino flavors (tracks vs. cascades). However, in the low-energy regime of 5–100 GeV, particularly within the DeepCore sub-array, events often exhibit sparse photon detection (specifically $10 \le n_{hit} < 20$). In this regime, the light pattern provides weak constraints on event morphology, limiting the separation power of state-of-the-art graph neural networks (GNNs) like DynEdge. While GNNs utilize raw per-pulse information (position, time, charge), they normalize timestamps to the microsecond-scale event duration, potentially compressing the critical nanosecond-scale early-time structure that distinguishes muon tracks from electromagnetic cascades. The paper addresses the challenge of whether nanosecond-scale timing information on individual Digital Optical Modules (DOMs) carries flavor-identification information that is complementary to existing morphology-based classifiers.

Methodology
The authors introduce WavePID, a compact, template-based log-likelihood-ratio classifier designed to exploit nanosecond-scale timing on individual DOMs. The methodology proceeds in three stages:

1. Microphysics Motivation (Geant4 Simulation): Using the Geant4 framework and the IceCube OMSim model, the authors simulated photon arrival time distributions (PATDs) for 5–100 GeV muons and electrons in ice. They found that muon tracks produce a sharp, direct Cherenkov front, resulting in a fast-rising PATD dominated by primary particle emission. Conversely, electromagnetic cascades (electrons) produce light via Bremsstrahlung secondaries over several nanoseconds, resulting in a broader PATD. This difference is most pronounced in the first $\sim10$ ns after the first photon arrival.
2. Observable Construction: WavePID operates on reconstructed ATWD pulses from individual DOMs. For each DOM $n$, it constructs a feature vector $\mathbf{x}_n$ comprising three physics-motivated observables:
   • $r_n$: The distance from the DOM to the reconstructed interaction vertex.
   • $\xi^t_n$: The early-charge fraction, defined as the fraction of total charge arriving within the first $t$ nanoseconds (specifically $t=14$ ns) after the first pulse.
   • $\Delta t_n$: The time of the first pulse on the DOM relative to the earliest pulse on any DOM in the event.
3. Classification Algorithm: The feature space is discretized into 3D bins. The algorithm constructs shape-only templates (probability mass functions) for "track" and "cascade" hypotheses using labeled Monte Carlo (MC) simulations. The test statistic (TS) is a binned log-likelihood ratio:
   $$TS_e = \log \frac{\mathcal{L}_{cascade}}{\mathcal{L}_{track}} = \sum_{i,j,k} H^e_{ijk} \log \frac{\lambda^{cascade}_{ijk}}{\lambda^{track}_{ijk}}$$
   where $H^e_{ijk}$ is the observed occupancy of DOMs in a specific bin for event $e$, and $\lambda$ represents the template probabilities.

Key Results
The performance of WavePID was evaluated on IceCube DeepCore Monte Carlo ensembles representing dominant detector and ice-model uncertainties (31 systematic variations). The study focused on a sample enriched in cascades (selected by a DynEdge score $\le 0.4$) with hit multiplicities between 10 and 20.

• Performance Improvement: In the sparse hit regime ($10 \le n_{hit} < 20$), WavePID achieved a higher Area Under the Curve (AUC) of 0.571 compared to 0.542 for DynEdge alone within the same pre-selected sample. This demonstrates that the early-time pulse structure encodes flavor information not captured by the GNN.
• Purity Gain: Applying a threshold to the WavePID test statistic increased the cascade purity. At an operating point retaining 20% of cascades, the purity increased from a baseline of 0.66 (DynEdge only) to 0.75. This gain of $+7\% \pm 1\%$ remained stable across all 31 detector systematic ensembles, indicating robustness to uncertainties in DOM efficiency, hole ice, and bulk ice optical properties.
• Systematics: The separation power remained stable across variations in detector parameters, confirming that the signal is driven by physical timing differences rather than instrumental artifacts.

Significance and Claims
The paper claims that WavePID demonstrates the viability of using nanosecond-scale single-PMT timing as a first-principles-motivated observable for flavor identification in neutrino telescopes. The authors emphasize that this is, to their knowledge, the first explicit use of single-DOM nanosecond timing for particle identification in a neutrino telescope.

The significance lies in the complementarity of the method: WavePID does not replace morphology-based classifiers but adds a new dimension of information. The discrimination is concentrated in the regime where the primary lepton's leading Cherenkov front is resolved by per-DOM pulse timing, a feature that existing classifiers may miss due to temporal normalization.

The authors conclude that these results motivate the exploitation of fine-grained per-module timing in future low-energy classifiers and in the design of next-generation neutrino telescopes (such as IceCube-Gen2, KM3NeT/ORCA, and others) that utilize multi-PMT or hybrid optical modules with improved timing resolution. The paper notes that the specific configuration presented is a baseline implementation, and the optimal choice of observables and binning depends on specific physics targets. Application to real IceCube data and quantification of physics impact are reserved for future dedicated studies.