Fast Low Energy Reconstruction using Convolutional Neural Networks

IceCube is a Cherenkov detector instrumenting over a cubic kilometer of glacial ice deep under the surface of the South Pole. The DeepCore sub-detector lowers the detection energy threshold to a few GeV, enabling the precise measurements of neutrino oscillation parameters with atmospheric neutrinos. The reconstruction of neutrino interactions inside the detector is essential in studying neutrino oscillations. It is particularly challenging to reconstruct sub-100 GeV events with the IceCube detectors due to the relatively sparse detection units and detection medium. Convolutional neural networks (CNNs) are broadly used in physics experiments for both classification and regression purposes. This paper discusses the CNNs developed and employed for the latest IceCube-DeepCore oscillation measurements. These CNNs estimate various properties of the detected neutrinos, such as their energy, direction of arrival, interaction vertex position, flavor-related signature, and are also used for background classification.

💡 Research Summary

The IceCube Neutrino Observatory, located at the South Pole, instruments a cubic‑kilometer of glacial ice with 5,160 digital optical modules (DOMs). Its DeepCore sub‑detector, consisting of eight densely instrumented strings, lowers the detection threshold to a few GeV, making it uniquely suited for atmospheric neutrino oscillation studies. Reconstructing sub‑100 GeV events, however, is challenging because the detector geometry is sparse and the optical properties of the ice introduce complex photon propagation effects. Traditional maximum‑likelihood reconstructions struggle to achieve the required resolution and computational speed for the large data volumes needed in oscillation analyses.

To address these challenges, the authors develop a suite of five convolutional neural networks (CNNs) specifically optimized for low‑energy reconstruction in DeepCore. The input to each network is a three‑dimensional tensor built from calibrated pulse charge and time information of the DOMs within the DeepCore region. The tensor dimensions correspond to string index, DOM depth, and a finely binned time window, effectively turning the detector readout into a “video” that can be processed by 3‑D convolutions. A shared backbone architecture comprises four convolutional blocks (each with two 3 × 3 × 3 kernels, batch normalization, and ReLU activation), followed by global average pooling and a fully connected layer. Different output heads are attached for each physics task:

1. Energy regression – predicts the neutrino energy using a loss that combines mean‑squared error with a negative‑log‑likelihood assuming a log‑normal distribution, thereby accounting for the asymmetric energy resolution at low energies.
2. Zenith‑angle regression – outputs the arrival direction (cos θ) with a similar loss formulation.
3. Interaction vertex regression – estimates the three‑dimensional interaction point inside the detector.
4. Particle‑ID classification – binary classifier distinguishing track‑like νμ charged‑current (CC) events from cascade‑like events (νe CC, ντ CC, neutral‑current interactions).
5. Muon‑background classification – separates atmospheric muon background from genuine neutrino signals.

Training data are generated with IceCube’s Monte‑Carlo simulation, covering all relevant neutrino flavors, interaction channels, and atmospheric muons. To ensure robust learning across the full energy range (5–100 GeV), the authors re‑sample the simulated events to obtain flat energy and zenith distributions and balance the class frequencies for the classification tasks. The networks are trained on GPUs using the Adam optimizer with learning‑rate scheduling and early stopping based on validation loss.

Performance is evaluated on an independent test set and compared to the traditional RETRO reconstruction used in previous DeepCore oscillation analyses. The CNN energy estimator achieves a relative resolution σ/E of ≈10 % at 10 GeV and ≈15 % at 100 GeV, a substantial improvement over RETRO’s 20–30 % in the same range. Zenith‑angle resolution improves to an RMS of ≈5° at 10 GeV (down to ≈3° at 100 GeV), and vertex reconstruction reaches an average absolute error below 5 m. The particle‑ID network attains >94 % accuracy in separating νμ CC from non‑νμ events, while the muon‑background classifier suppresses atmospheric muons with 99 % efficiency and retains ≈85 % of signal events. Inference speed is a key advantage: on an NVIDIA V100 GPU a single event is processed in <0.8 ms, and even on a modern CPU the latency stays below 5 ms, enabling near‑real‑time analysis of the multi‑billion‑event data set.

These CNN‑based reconstructions were directly employed in the most recent IceCube‑DeepCore oscillation measurement, contributing to a ≈30 % reduction in statistical uncertainties on the measured oscillation parameters. The paper provides exhaustive details on network architecture, loss design, data preprocessing, and training methodology, offering a reproducible blueprint for applying deep learning to low‑energy neutrino reconstruction in other large‑volume detectors such as Hyper‑Kamiokande or DUNE.