Learning Efficient Representations of Neutrino Telescope Events

TL;DR

Neutrino telescope data are large, high-dimensional, and sparsely structured due to photon arrival time distributions (PATDs). The paper introduces om2vec, a transformer-based variational autoencoder that maps per-OM PATDs into compact latent representations, enabling efficient reconstruction and downstream analyses such as angular reconstruction. Results show that latent representations retain essential information and achieve comparable performance to full-timing inputs while delivering substantial speedups and enabling image-like ML approaches on latent data. The approach is validated on Prometheus-simulated IceCube-like data, with implications for reduced data throughput and real-time analysis; code and datasets are available on GitHub.

Abstract

Neutrino telescopes detect rare interactions of particles produced in some of the most extreme environments in the Universe. This is accomplished by instrumenting a cubic-kilometer scale volume of naturally occurring transparent medium with light sensors. Given their substantial size and the high frequency of background interactions, these telescopes amass an enormous quantity of large variance, high-dimensional data. These attributes create substantial challenges for analyzing and reconstructing interactions, particularly when utilizing machine learning (ML) techniques. In this paper, we present a novel approach, called om2vec, that employs transformer-based variational autoencoders to efficiently represent the detected photon arrival time distributions of neutrino telescope events by learning compact and descriptive latent representations. We demonstrate that these latent representations offer enhanced flexibility and improved computational efficiency, thereby facilitating downstream tasks in data analysis.

Figures (6)

• Figure 1: An artistic rendition of a cascade-like event, showcasing how photon arrival time distributions (PATDs) are recorded in neutrino telescopes. A neutrino interaction producing photons occurs in the detector medium, where it is surrounded by photon-detecting optical modules (OMs). The photon arrival times are recorded and counted, as shown in the histograms in the above figure. The amount of photons a particular OM sees is highly variable and generally depends on its proximity to the interaction point. The main goal of om2vec is to convert the PATD on each OM in the event into a fixed-size latent representation.
• Figure 2: Model architecture of om2vec. Input embeddings are generated from the binned PATDs. Each encoder (decoder) block operates on both the feature and sequence dimensions, where feedforward layers are used to downsample (upsample) the length of the distributions. In the decoder, a memory embedding is learned to keep the decoder independent of the encoder.
• Figure 3: Jensen-Shannon distance between the true input PATD and the reconstructed PATD across different methods, plotted as a function of the number of detected photons in the input PATD. Greater values of the JS distance implies a worse fit. In the AGMM methods, the sudden jump is caused by the fit only being performed when the number of photon hits is greater than the number of Gaussian components. The lower panel illustrates the percentage of PATDs in each number of photons bin, relative to the total dataset.
• Figure 4: Failure percentage ($>0.99$ JS distance) as a function of the number of photons on the input PATD. om2vec is always at 0% in this figure.
• Figure 5: Two example PATDs, the bottom one being a "double-bang" signature, characterized by two distinct peaks. om2vec is able to reconstruct both peaks, while the AGMM is noticeably less accurate.