Skip to content

Lightning Tracks Event Selection

This section documents the Lightning Tracks (LT) event selection on a conceptual level: selection steps, final sample performance, which decisions went into the cuts, and how the underlying machine-learning models were trained and validated. The goal is that someone reading this page understands what the selection is and why it looks the way it does, without needing to care about technical implementation details.

The technical details of the processing workflow itself (dataset management, DAG creation, containers, etc.) are documented separately in Processing.

Design Goals

With the ultimate goal of improving IceCube’s sensitivity to neutrino sources, the Lightning Tracks selection is built around a few core ideas:

  • Signal capture: retain high effective area for track-like neutrino events over a wide energy range by eliminating the need for computationally cheap early cuts (e.g. on total visible charge). Instead of such simple cuts, we use more sophisticated, yet relatively cheap-to-apply, machine-learning (ML) filter models as the very first step. This also allows us to skip intermediate selection levels at medium–high data rates where we would still have to rely on simple reconstructions for further cuts due to computational constraints. In other words: do not cut too hard too early, since any events removed early on cannot be recovered later.
  • Background rejection: suppress atmospheric muons and poorly reconstructed events, but only to the degree to which it improves point-source sensitivity while achieving at least reasonable data–MC agreement. In other words, do not cut too hard later on either.
  • Angular performance: achieve competitive or better angular resolution compared to previous track analyses by
    • deploying new and improved ML reconstruction algorithms;
    • eliminating the need to leverage cheaper but worse-performing algorithms or settings at intermediate selection levels by skipping those levels entirely, as mentioned above. For example, an event that can be reconstructed with high accuracy by one of our best but most computationally expensive reconstruction algorithms might not be reconstructible by a cheaper alternative. As a result, it might previously have been removed in intermediate selection levels of a conventional multi-level event selection.
  • Prioritize sensitivity: the primary limiting factor when it comes to developing a new event selection in IceCube is human resources. There are practically infinitely many sub-problems that can be optimized, but only a few have a significant impact on the quality of the final science results. This leads to marginal returns on time investment when working on any less impactful optimization problem. We focus on those problems for which we see the largest potential for sensitivity improvements, while disregarding others where possible. For example, early sensitivity testing showed that energy resolution does not meaningfully impact sensitivity to neutrino sources, and thus the decision was made to use the well-established and robust MuEX energy estimator instead of investing significant time into exploring improved energy reconstruction.

Throughout this document we refer to two main physics samples:

  • Starting tracks: track-like events whose neutrino interaction vertex is contained within the convex hull defined by the detector’s optical modules. In the context of Lightning Tracks, this subset is referred to as Starting Lightning Tracks (SLT).

  • Throughgoing tracks: track-like events that originate outside the instrumented volume but traverse the detector’s convex hull. Within Lightning Tracks, this subset is denoted as Throughgoing Lightning Tracks (TLT).

Both channels are treated as separate sub-selections with a hierarchical structure: events that pass the initial starting event filter are not considered in the throughgoing sub-selection. Both sub-selections run alongside each other during the processing. This way we automatically create two disjoint final sub-samples that can be trivially combined at analysis level to yield the full sample without having to account for overlap. Treating the topologies as separate samples in Csky is crucial for optimal sensitivity because the signal acceptance and background PDFs differ significantly.

High-Level Selection Flow

At a conceptual level, the LT selection proceeds in four stages:

  1. Filter models (Lightning CNN/MLP filters)
  2. Event reconstruction (direction, energy, vetoes)
  3. Final quality cuts and MLP-based final starting/throughgoing selection
  4. High-level export for physics analysis (HDF → NumPy → Csky)

Only the filtering step removes events from the i3 files. All filtered events are reconstructed with the final best-performing reconstruction algorithms. Events passing the final cuts are only labeled, not removed from i3. This allows the filtered sample to be used as a potential base for other use cases (e.g. a diffuse analysis).

1. Filter Models (LT Filters)

The selection begins with machine-learning–based filters that operate directly on Level 2 events:

  • LCSC starting-track CNN
    • A convolutional neural network (CNN) using DOM-level summary features as input to identify starting-track-like events purely based on topology.
  • LCSC upgoing-track CNN
    • A CNN using DOM-level summary features as input to identify upgoing-track-like events purely based on topology.
  • Downgoing-throughgoing MLP
    • A multilayer perceptron (MLP) using tabular (event-level) Level 2 variables as input to identify downgoing-throughgoing-track-like events with a relatively high likelihood (“astroness”) of being single muons from astrophysical neutrino interactions based on prior knowledge and assumptions about astrophysical signal and atmospheric background distributions.

Each event receives one score from each filter, stored under a common naming scheme (e.g. LCSC_model_predictions and related keys). Simple score thresholds define an initial candidate pool. Any event passing at least one of the following three cuts is included in our filter level, which is then fully reconstructed and saved:

  • Starting candidates: model score cut of 0.999.
  • Upgoing candidates: model score cut of 0.99.
  • Throughgoing-downgoing candidates: model score cut of 0.99.

Those cuts result in a data rate of roughly 20 mHz for starting and 70 mHz for throughgoing events when applied to Level 2 data (a detailed season-by-season breakdown is given in Experimental Data). This corresponds to a rate reduction of 99.98%, sufficient to allow computationally expensive reconstructions to be run on all surviving events. The filter models themselves, including their training and the choice of cut thresholds, are described in detail in Final Cut Models.

2. Reconstructions

For all filter-passing candidates, several dedicated reconstruction tools are run to obtain the observables used in the final selection:

  • Gerrit’s RNN directional reconstruction

    • Provides a fast, full track reconstruction (direction, location, time, uncertainty) implemented as a convolutional recurrent neural network (CRNN).
    • Used as one of the two main direction estimates for quality cuts and as track seed for MuEX. It is not used for final directional reconstruction on analysis level.
    • Unfortunately, Gerrit’s PhD thesis is not published yet and there is no official documentation write-up until then. You can find more information in some of Gerrit’s past presentations (e.g. here). It was also independently verified using the moon shadow.

  • Gerrit’s RNN energy reconstruction

    • Same CRNN architecture, but trained to predict muon energy losses.
    • Originally we planned to use it for energy and vertex reconstruction, but it is not currently utilized in LT after all and can safely be excluded from the processing if computing resources are tight. However, so far it was included in all processing so it could be helpful if someone wanted to use it in future work without having to re-run the processing.

  • MuEX energy estimator

    • Traditional energy reconstruction included in IceTray based on the observed charge along the muon track.
    • We use the RNN reconstructed track as input (since TNF does not reconstruct a full track including location and timing, only the direction in zenith and azimuth).
    • Serves as the nominal reconstructed energy for cuts and physics analysis.

  • ESTES Starting Track Veto (STV)

    • Conventional (no ML) muon veto algorithm included in IceTray.
    • Applied only to starting-track candidates.
    • Run twice: with and without stochastic suppression. The main variable utilized in LT is the miss probability \(p_\text{miss}\), used as one of the input features for the final starting-cut model.
      • A high miss probability indicates the event is likely an entering muon. Find more on the ESTES Wiki.

  • Thorsten’s Normalizing Flow (TNF) directional reconstruction

    • Provides a powerful directional fit with extractable angular uncertainty by estimating the entire likelihood space of the event in azimuth and zenith from DOM-level features via normalizing flow.
    • We extract both a best-fit direction and an event-by-event error estimate, which are used in the final cuts and as the final directional reconstruction and angular error for neutrino source analyses with Csky.
    • You can learn more about the specifics of the method in Thorsten’s material and more about how we use it for angular error modeling with Csky in the dedicated Angular Error Calibration section.

This combination ensures that each candidate has at least two independent direction estimates (RNN and TNF), a robust energy estimate (MuEX), and, for starting tracks, a veto-based containment variable (STV’s \(p_\text{miss}\)). In combination with the filtering model scores those variables are used for the final sample cuts.

3. Final Quality Cuts and MLP Selections

The final selection is defined by a set of quality cuts and two small MLPs that map the full information into final signal likelihood variables for starting and throughgoing tracks.

3.1 Baseline Quality Cuts

Quality cuts act as a first pass to remove clearly poorly reconstructed events and improve data–MC agreement. The quality cuts are:

  • Directional consistency
    • Maximum angular error estimates from both RNN and TNF reconstructions:
      • \(\sigma_{\text{RNN}} < 5^\circ\)
      • \(\sigma_{\text{TNF}} < 5^\circ\)
    • Maximum separation between RNN and TNF best-fit directions:
      • \(\Delta\psi(\hat{n}_{\text{RNN}}, \hat{n}_{\text{TNF}}) < 5^\circ\)
      • We call this the Reco Separation cut. It significantly helps with removing misreconstructed events and eliminates the need for further topology classification (e.g. treating cascades explicitly). It is based on a simple idea: when the reconstructions fail (in terms of poor accuracy and poor error estimation), they at least do not fail in the same way, since they are completely different algorithms based on different data, resulting in unique failure cases. The most likely cause for such failure cases are events not at all or not sufficiently covered by the training data, e.g. cascades or rare coincident events.
  • Charge and energy requirements
    • Minimum homogenized total charge (a proxy for brightness and reconstruction stability):
    • Minimum reconstructed MuEX energy:
      • \(E_{\text{MuEX}} > 10\,\text{GeV}\).

These numbers can be refined per analysis and are ultimately tuned to achieve the desired balance between efficiency, background rejection, and systematic robustness.

3.2 Final MLP Models

After quality cuts, lightweight MLPs define the final starting and throughgoing selections:

  • final_starting_cut: MLP trained to distinguish starting tracks from atmospheric background (entering muons).
  • final_throughgoing_cut: MLP trained to select clean throughgoing track candidates with a high likelihood of astrophysical origin.

Both models take as input a compact set of features derived from:

  • LCSC filter scores.
  • RNN and TNF reconstruction outputs (zenith, uncertainties).
  • Reco Separation.
  • MuEX energy and homogenized total charge.
  • STV miss probabilities (with and without stochastic suppression) in the case of starting tracks.

In the default configuration, the MLPs are not used to immediately discard events. Instead, they assign a probability-like score that is stored for each event, and the final sample is selected by applying thresholds at the notebook/postprocessing stage. This keeps the processing pipeline maximally flexible: different physics analyses can retune the thresholds or work with continuous weights. See Sensitivity Optimization for how this can be used to optimize the sensitivity for different kinds of neutrino source analyses.

4. High-Level Analysis Products

Events passing the filter and reconstruction chain are exported into high-level formats:

  1. Per-run HDF5: export tabular data from each i3 file.
  2. Merged NumPy arrays: merge HDF5 files into per-dataset tables.
  3. Csky-compatible NPY: final sample mapped to Csky’s expected fields, with optional overlap removal against other samples (e.g. DNNC).