Deep Learning Framework for Enhanced Neutrino Reconstruction of Single-line Events in the ANTARES Telescope

TL;DR

The paper presents N-fit, a neural-network-based reconstruction framework for single-line neutrino events in ANTARES that integrates deep convolutional networks, mixture density networks, and transfer learning to separately handle track and shower topologies. By decomposing tasks into direction, distance to the detector, energy, and a TL-driven classifier, N-fit achieves significant improvements in angular resolution (notably enabling azimuthal information) and energy estimation, with robust uncertainty quantification. Transfer learning, including PCA-based knowledge distillation for energy and direct TL for classification, yields better energy predictions and higher classifier performance, validated on MC and real data with stability under robustness tests. The approach demonstrates practical gains for multimessenger astrophysics and can be extended to KM3NeT and other detectors, offering a scalable path to enhanced physics analyses with single-line events.

Abstract

We present the $N$-fit algorithm designed to improve the reconstruction of neutrino events detected by a single line of the ANTARES underwater telescope, usually associated with low energy neutrino events ($\sim$ 100 GeV). $N$-Fit is a neural network model that relies on deep learning and combines several advanced techniques in machine learning --deep convolutional layers, mixture density output layers, and transfer learning. This framework divides the reconstruction process into two dedicated branches for each neutrino event topology --tracks and showers-- composed of sub-models for spatial estimation --direction and position-- and energy inference, which later on are combined for event classification. Regarding the direction of single-line events, the $N$-Fit algorithm significantly refines the estimation of the zenithal angle, and delivers reliable azimuthal angle predictions that were previously unattainable with traditional $χ^2$-fit methods. Improving on energy estimation of single-line events is a tall order; $N$-Fit benefits from transfer learning to efficiently integrate key characteristics, such as the estimation of the closest distance from the event to the detector. $N$-Fit also takes advantage from transfer learning in event topology classification by freezing convolutional layers of the pretrained branches. Tests on Monte Carlo simulations and data demonstrate a significant reduction in mean and median absolute errors across all reconstructed parameters. The improvements achieved by $N$-Fit highlight its potential for advancing multimessenger astrophysics and enhancing our ability to probe fundamental physics beyond the Standard Model using single-line events from ANTARES data.

Figures (23)

• Figure 1: Schematic representation of the angles $\theta$ and $\phi$ that define the direction of a neutrino, $\nu$, in the ANTARES coordinate system.
• Figure 2: Schematic diagram of a simple feed-forward neural network. Circles represent neurons, and arrows represent synaptic connections. In general, more than a single output can be considered.
• Figure 3: Example of a normalized RGB image from a track-like event ($\hbox{$\tiny{\hbox{( )}}$}\bar{\nu}_\mu^{\mathrm{CC}}$).
• Figure 4: Details of the direction neural network architecture. Note that for the $\theta$ prediction, we scaled the hyperbolic tangent activation function for $\mu_\theta$, so that its values lay in $[0, \pi]$ radians. No scaling was necessary for the prediction of $\phi$ since its estimation was derived from the Cartesian $\{X,Y\}$ components of the unit vector.
• Figure 5: The error on angles $\theta$ (left) and $\phi$ (right) as a function of the predicted uncertainty for the track branch. The mean and median error stay close to zero with no significant bias. The first and third quartiles behave as expected for a Gaussian distribution, especially for $\theta$.