Neutrino Telescope Event Classification on Quantum Computers

Abstract

Quantum computers represent a new computational paradigm with steadily improving hardware capabilities. In this article, we present the first study exploring how current quantum computers can be used to classify different neutrino event types observed in neutrino telescopes. We investigate two quantum machine learning approaches, Neural Projected Quantum Kernels (NPQKs) and Quantum Convolutional Neural Networks (QCNNs), and find that both achieve classification performance comparable to classical machine learning methods across a wide energy range. By introducing a moment-of-inertia-based encoding scheme and a novel preprocessing approach, we enable efficient and scalable learning with large neutrino astronomy datasets. Tested on both simulators and the IBM Strasbourg quantum processor, the NPQK achieves a testing accuracy near 80%, with robust results above 1 TeV and close agreement between simulation and hardware performance. A simulated QCNN achieves a ~70% accuracy over the same energy range. These results underscore the promise of quantum machine learning for neutrino astronomy, paving the way for future advances as quantum hardware matures.

Figures (5)

• Figure 1: Moment of inertia representations and event histogram binned in energy decade On the left, the two data categories of interest, (a) tracks and (b) cascades, are shown. Schematic representations of the typical shapes are superimposed on each, illustrating the rationale behind employing moments of inertia as a distinguishing feature. In the event displays, each circle represents an Optical Module (OM): the size of the circle indicates the number of photons detected by that OM (larger circles correspond to more photons), and the color shows the photon arrival time (with red for earlier and green for later arrivals). On the right (c), the 2D histogram displays the number of photons per event, grouped by energy bins. For each energy decade, 1000 events were generated — half are tracks and half are cascades.
• Figure 2: Distribution of encoded variables. Each column represents a different energy range from 100 GeV to 1 PeV in four decades. Lighter dots corresponds to Track events, while darker dots are Cascade events. Top row shows $I_1-I_0$ as a function of $I_2-I_1$, while bottom row shows CoM as a function of these two variables. We display 300 data points for clarity. In our analysis, we exploit the separation shown in the bottom figures.
• Figure 3: Schematics of quantum architectures. (a) Neural projected quantum kernel construction. The embedding is constructed using the optimal parameters $\boldsymbol{\theta}$ obtained from training a QNN. The reduced density matrices of the first qubit are stored on a classical computer and used to construct the kernel entries $k_{ij}$, which are subsequently fed into a SVM for prediction. (b) Quantum convolutional neural network. Input data is amplitude-encoded and processed through convolutional layers consisting of parameterized unitaries. Pooling is performed via weighted measurements, effectively halving the number of qubits at each step. A final fully connected layer precedes measurement, enabling parameter optimization through a classical optimizer.
• Figure 4: Classification results for tracks vs. cascades across different energy ranges using different feature sets. (a) Training and test accuracies using the three raw moments of inertia as input features, $\vec{x} = (I_2, I_1, I_0)$. (b) Results using a feature set composed of two moment differences and the center-of-mass, $\vec{x} = (I_2 - I_1, I_1 - I_0, \texttt{CoM})$. For both feature sets, models are trained on 200 events and tested on 120. Metrics are reported as the mean and standard deviation across 5 random dataset splits. Results include: neural projected quantum kernels (NPQK) evaluated both in ideal simulation and on IBM's Strasbourg QPU, and a quantum convolutional neural network (QCNN) evaluated under ideal simulation. (c) Visualization of the 40 qubit pairs we used, shown in blue, selected from the 127-qubit IBM Strasbourg quantum device. The pairs were chosen to maximize physical separation, enabling the simultaneous sampling of 40 reduced density matrices and thereby reducing the overall QPU runtime.
• Figure 5: Classification results (F1 score) under class imbalance. Both training and test F1 scores are shown. The training was performed on a balanced dataset, while the test set featured a 90--10 imbalance in favor of tracks. As before, 200 training points and 120 test points were used. Results correspond to the feature set composed of two moment differences and the center of mass, $\vec{x} = (I_2 - I_1, I_1 - I_0, \texttt{CoM})$, within the energy range of 1--10 TeV. Shown are the results for classical algorithms--Random Forest (RF) and Support Vector Classifier with Gaussian kernel (SVC)--alongside those obtained with the neural projected quantum kernel, both in ideal simulation and on real quantum hardware (IBM Strasbourg and IBM Basque Country).