ANTS: Advanced Northern Tracks Selection

• The paper presents a graph-based framework that combines a transformer autoencoder with 70-layer GCNs to boost neutrino event selection efficiency in IceCube.
• It processes raw DOM-level Cherenkov data, achieving approximately 25% improvement in statistics over traditional BDT methods at 99.8% purity.
• ANTS extends its utility beyond binary classification to energy regression and event topology, offering potential for real-time neutrino filtering.

Searching arXiv for the cited papers to ground the article and clarify the acronym usage. arxiv_search(query="^^^^^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^^^^^"Advanced Northern Tracks Selection\"^^^^^^^^2 OR \2^^^^^^^^"Trade-offs between Selection Complexity and Performance when Searching the Plane without Communication\"", max_results=^^^^2 OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^) Advanced Northern Tracks Selection (ANTS) is an event-selection framework in the IceCube Neutrino Observatory that employs a graph convolutional neural network to distinguish between muons induced by neutrinos and those generated by cosmic-ray air showers. The method is designed to use both sensor data and the geometric arrangement of the detector’s photomultiplier tubes (PMTs): each module is represented as a node in a graph, and the learned representation integrates local and global detector information for muon-track selection in the northern sky (&&&^^^^2 OR \2^^^^&&&).

^^^^2 OR \2^^^^. Experimental setting and selection objective

IceCube is a cubic-kilometer detector located in the Antarctic ice at the geographic South Pole, and it reads out over 5,^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^ PMTs to detect Cherenkov light produced by secondary particles. Within this instrument, a central analysis problem is to distinguish up-going muons from neutrino interactions from down-going atmospheric muons. ANTS is positioned as a replacement for, or extension of, the standard Northern Tracks selection, which uses a series of quality cuts and two Boosted Decision Trees (BDTs) relying on high-level reconstruction features for background rejection, achieving approximately 99.8% purity but potentially discarding useful events (&&&^^^^2 OR \2^^^^&&&).

The defining methodological shift is from handcrafted, reconstruction-level observables to learned representations constructed directly from detector-level activity. This design choice is significant because the detector is inherently irregular and three-dimensional: PMTs are embedded in the ice in a fixed geometry, and the event signature is distributed across both space and time. A graph formulation allows the selection to encode this non-Euclidean structure directly rather than approximating it through engineered summary variables.

2. DOM-level representation and preprocessing

ANTS begins from DOM-level data. Each Digital Optical Module records a time series of Cherenkov photon hits expressed as charge-time pairs. Preprocessing includes merging close charge-time pairs within ^^^^2 OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^ ns for each DOM to reduce digitization artifacts, normalizing features by standard deviation, subtracting the per-DOM median time, applying a signed square root transformation for dynamic range, and shifting the data so the first hit is at PRESERVED_PLACEHOLDER_^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^. Relative time information between different DOMs is also retained as auxiliary information for higher-level pattern recognition (&&&^^^^2 OR \2^^^^&&&).

A transformer-based autoencoder is then used to learn a compressed DOM-wise representation from these raw time series. The input is each DOM’s merged and normalized time-charge sequence, padded or masked to 256 hits. The encoder consists of four multi-head attention layers, each with 4 heads, followed by global average pooling and an MLP, yielding ^^^^2 OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^ latent features per DOM. The decoder mirrors the encoder, and training uses mean squared error loss for reconstruction. The reported outcome is that the latent features represent each DOM’s photon timing and charge pattern without loss of signal class information and are robust between data and simulation (&&&^^^^2 OR \2^^^^&&&).

This representation stage is central to the ANTS design because it separates low-level waveform compression from event-level classification. A plausible implication is that the later graph network can focus on inter-DOM structure rather than relearning the internal temporal regularities of each DOM from scratch.

3. Graph construction and network architecture

After latent feature extraction, the detector is represented as a graph. Each DOM is a node, and edges connect each node to its ^^^^2 OR \2^^^^6 nearest neighbor DOMs according to detector geometry; edge weights encode proximity. The node feature vector is ^^^^2 OR \2^^^^6-dimensional, combining the autoencoder output with auxiliary information. The classifier then applies 7^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^ graph convolutional network layers embedded in ResNet blocks, followed by partition pooling, global average pooling, and small multi-layer perceptrons for the final decision boundary (&&&^^^^2 OR \2^^^^&&&).

In simplified form, for an event PRESERVED_PLACEHOLDER_^^^^2 OR \2^^^^ with node features $X=\{x_1,\ldots,x_N\}$, the graph is $G=(V,E)$, and the layer update is written as

$$h^{(l+1)}_i = \sigma \left( \sum_{j \in \mathcal{N}(i)} w_{ij} \cdot W^{(l)} h^{(l)}_j + b^{(l)} \right),$$

where $h^{(0)}_i=x_i$, $\mathcal{N}(i)$ denotes the neighbors of node $i$, and $W^{(l)}$ and $b^{(l)}$ are learnable parameters (&&&^^^^2 OR \2^^^^&&&).

The primary output is a binary classifier for signal versus background, but the same architecture is described as extensible to regression, such as energy estimation, and to multi-class classification, such as event-topology inference. This architectural generality suggests that ANTS is not merely a fixed cut replacement; it is a graph-based event representation and inference stack for multiple downstream IceCube tasks.

4. Comparative performance and operating point

The reported performance gains are expressed most directly through background rejection and signal retention. On the ROC curve, the BDT baseline reaches an AUC of ^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^.9833, whereas the ANTS GCN reaches ^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^.9992. At matched purity of 99.8%, ANTS retains approximately 95% of true neutrino events, corresponding to approximately 25% improvement in statistics over the BDT baseline. In ^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^.23 yr of data, the event yield is approximately 23,^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^ for ANTS versus approximately ^^^^2 OR \2^^^^8,5^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^ for the BDT selection (&&&^^^^2 OR \2^^^^&&&).

The gains are especially pronounced near the horizon, described as the most background-prone region, while maintaining high efficiency across the up-going muon sample. In operational terms, this means that ANTS improves the selection precisely where atmospheric contamination is most challenging.

Aspect	BDT Baseline	ANTS
Input features	High-level reconstruction	Raw, DOM-based (+ learned)
Architecture	2x Boosted Decision Trees	Transformer-Autoencoder + GCN
AUC	^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^.9833	^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^.9992
Signal efficiency at 99.8% purity	—	~95% retained
Event recovery in ^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^.23 yr	~^^^^2 OR \2^^^^8,5^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^	~23,^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^
Extensions	Cascade only	Energy regression, topology

These results indicate that the principal advantage of ANTS is not only higher aggregate classification quality, but improved efficiency at a fixed high-purity operating point. For IceCube analyses that are statistics-limited after strong background rejection, that distinction is consequential.

5. Validation, scientific significance, and deployment scope

The reported validation emphasizes excellent data/MC agreement across all relevant variables. The same architecture is also stated to perform strongly for energy regression, reconstructing deposited energy, and for event-topology classification with low misclassification rates. The framework is therefore presented as both a selection algorithm and a reusable detector-learning architecture (&&&^^^^2 OR \2^^^^&&&).

The scientific significance is tied directly to event statistics and downstream analysis sensitivity. The improved efficiency is described as equivalent to years of additional detector exposure without new infrastructure. More high-purity events improve the statistical precision of the astrophysical neutrino spectrum, source searches including NGC ^^^^2 OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^68, and atmospheric studies. The work also notes that the fast inference speed creates the possibility of deployment earlier in the data chain, including real-time filtering and potentially the real-time trigger; domain adaptation to reduce data/simulation mismatches is also identified as a possible extension (&&&^^^^2 OR \2^^^^&&&).

These points situate ANTS within a broader transition in neutrino astronomy from reconstruction-centric event selection to end-to-end learned representations over detector graphs. A plausible implication is that similar architectures may increasingly unify filtering, classification, and regression tasks within a common graph-based framework.

6. Nomenclature and distinct uses of “ANTS” in the literature

The acronym “ANTS” appears in arXiv literature outside IceCube, and those usages should not be conflated with Advanced Northern Tracks Selection. One distinct example is the theoretical distributed-search work “Trade-offs between Selection Complexity and Performance when Searching the Plane without Communication,” which studies a group of agents collaboratively searching for a target in a two-dimensional plane and introduces the selection complexity metric PRESERVED_PLACEHOLDER_^^^^2 OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^. In that setting, PRESERVED_PLACEHOLDER_^^^^2 OR \2 OR \2^^^^ is identified as a crucial threshold for selection complexity, with upper and lower bounds on expected search time as a function of the number of agents PRESERVED_PLACEHOLDER_^^^^2 OR \22^^^^, target distance PRESERVED_PLACEHOLDER_^^^^2 OR \2^^^^3, memory PRESERVED_PLACEHOLDER_^^^^2 OR \2^^^^4, and probability granularity PRESERVED_PLACEHOLDER_^^^^2 OR \2^^^^5 (&&&^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^&&&).

Another distinct line of work is “Swarm behavior tracking based on a deep vision algorithm,” which addresses multi-ant tracking in videos. That framework follows a tracking-by-detection paradigm with a two-stage object detector using ResNet-5^^^^2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^, appearance descriptors learned by a separate ResNet, Kalman-filter motion prediction, and Hungarian assignment. It reports 95.7% mMOTA and 8^^^^2 OR \2^^^^.^^^^2 OR \2^^^^% mMOTP in indoor videos, 8^^^^2 OR \2^^^^.8% mMOTA and 8^^^^2 OR \2^^^^.9% mMOTP in outdoor videos, and a speed 6–^^^^2 OR \2id:(Lenzen et al., 2014) OR id:(Soldin et al., 10 Jul 2025) OR \2^^^^ times faster than existing methods for insect tracking (&&&^^^^2 OR \22^^^^&&&).

The overlap in terminology is therefore nominal rather than methodological. In current IceCube usage, ANTS denotes a graph-based northern-track event-selection framework for neutrino-induced muons; in other arXiv contexts, similar ant-related terminology refers either to decentralized search theory or to biological swarm tracking.