⚛️ phenomenology

Classical and Hybrid Quantum Machine Learning for Trigger-Like Event Selection on CMS Open Data: An Eight-Qubit, PCA-Constrained Benchmark

This study benchmarks four classical and four hybrid quantum machine learning models on an eight-qubit, PCA-constrained CMS open data trigger task, finding that while the classical artificial neural network outperforms all quantum counterparts in accuracy and ROC-AUC, trainable hybrid quantum embeddings (specifically the quantum convolutional network) achieve competitive results, serving as a controlled reference point rather than a demonstration of quantum advantage.

Original authors: Tariq Mahmood, Muhammad Awais Rafique, Talab Hussain, Juan Pablo Perez Aguilar, Alfredo Raya, Muhammad Ahsan

Published 2026-08-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tariq Mahmood, Muhammad Awais Rafique, Talab Hussain, Juan Pablo Perez Aguilar, Alfredo Raya, Muhammad Ahsan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, chaotic collisions inside the Large Hadron Collider, particles smash together at nearly the speed of light, creating a storm of debris. Most of this debris is ordinary background noise, but hidden within it are rare, precious events that could reveal new laws of nature. The challenge for physicists is not just to find these needles in a haystack, but to find them instantly. The machine produces data so fast that no computer could possibly save every single collision. Instead, a digital gatekeeper, called a trigger, must decide in a fraction of a second which events to keep and which to discard. This decision must be incredibly fast and incredibly accurate, or the rare discoveries will be lost forever. For years, scientists have used classical computer algorithms to make these choices, but a new question has emerged: could the strange rules of quantum mechanics offer a better way to sort this cosmic data?

A team of researchers set out to answer this by building a direct comparison between traditional computer models and a new generation of hybrid quantum models. They used real data from the Compact Muon Solenoid experiment at the collider, focusing on a specific task that mimics the trigger's job: identifying events where two particles combine to form a specific, short-lived particle known as the J/psi. To do this, they trained eight different computer brains. Four were built using standard, classical methods that run on today's supercomputers, while the other four were hybrid models that use a small quantum processor to help process the information. The researchers were careful to keep the playing field level, giving the quantum models a strict limit of eight quantum bits, or qubits, which is a realistic constraint for current technology. They also compressed the data down to its most essential features before feeding it into the quantum systems, ensuring a fair test of how well these emerging machines could learn.

The results showed that the most powerful classical model, a type of artificial neural network that mimics the way the human brain connects information, was the best performer overall. It correctly identified the target events with an accuracy of 93.53 percent and was able to rank the most likely candidates with a high degree of confidence. The other classical models, including those designed to recognize patterns in images or sequences, performed well but fell slightly behind. Among the quantum models, the most successful was a quantum version of a convolutional network, which reached an accuracy of 90.89 percent. This was a strong showing, coming very close to the performance of the best classical models. However, the other quantum approaches, including a quantum version of a memory-based network and a quantum support vector machine, struggled significantly, achieving much lower accuracy rates.

The study suggests that while quantum machines are not yet superior to their classical counterparts for this specific task, they are certainly capable of competing within a realistic budget. The researchers found that the quantum models that could adjust their internal settings through training performed much better than those with fixed settings. This indicates that the ability to learn and adapt is more important than the mere presence of quantum hardware. The quantum convolutional network, which uses a structure similar to how classical networks scan images, proved to be the most effective quantum approach, likely because its design helped it extract useful patterns from the compressed data more efficiently than the other quantum designs.

Despite these promising numbers, the authors are careful not to claim a victory for quantum computing. They point out that the task they set up was somewhat straightforward because the answer was already hidden in the data they provided; the models were essentially learning to reconstruct a known mathematical rule from the information given. This means the high scores do not necessarily prove that quantum computers can discover entirely new, complex patterns that classical machines cannot find. Furthermore, the quantum models were tested in a simulation rather than on actual quantum hardware, which means they did not face the noise and errors that real machines currently suffer from. The researchers also noted that the classical models had access to more raw information than the quantum models, which were forced to work with a compressed version of the data.

Ultimately, this work serves as a controlled benchmark rather than a final verdict. It establishes a clear reference point for how quantum and classical systems compare when they are asked to solve the same problem under the same constraints. The findings show that quantum models can reach a level of performance that is comparable to classical ones, but they have not yet surpassed them. For the future of particle physics, this means that while quantum machine learning is a promising tool worth pursuing, it is not yet a replacement for the classical algorithms that currently keep the world's most powerful particle collider running. The path forward involves refining these quantum models, testing them on real hardware, and finding ways to give them access to the same rich data as their classical rivals, all while keeping an eye on the speed and reliability required to catch the next great discovery in the making.

Technical Summary: Classical and Hybrid Quantum Machine Learning for Trigger-Like Event Selection

Problem Definition
The paper addresses the challenge of event triggering in high-energy physics, specifically within the context of the Compact Muon Solenoid (CMS) experiment at the Large Hadron Collider. The core problem is a binary classification task: distinguishing rare signal events from a vast background of collision data under strict latency and bandwidth constraints. The study utilizes CMS open data to construct a "trigger-like" scenario where the goal is to retain events falling within a specific invariant-mass window ( $3.0 \le M \le 3.2$ GeV), corresponding to the $J/\psi$ resonance, while discarding others.

Methodology
The authors conduct a controlled benchmark comparing four classical machine learning models against four hybrid quantum machine learning (QML) models. All models are trained on the same dataset (31,892 events) using a stratified 80/20 train/test split with a fixed decision threshold of 0.5.

Data Preprocessing: Inputs consist of reconstructed kinematic variables (four-vector components of two objects) and derived physics-motivated features: pseudorapidity difference ( $\Delta\eta$ ), wrapped azimuthal difference ( $\Delta\phi$ ), angular separation ( $\Delta R$ ), and total transverse momentum ( $p_{T,tot}$ ).
Quantum Constraints: All quantum models operate under a strict resource budget:
- Qubits: Fixed at 8.
- Feature Compression: Inputs are compressed via Principal Component Analysis (PCA) to 16 features before encoding.
- Simulation: All quantum circuits are executed using state-vector simulation (PennyLane lightning.qubit backend).

Model Architectures

Classical Baselines:
- Support Vector Machine (SVM): Uses a Radial Basis Function (RBF) kernel.
- Artificial Neural Network (ANN): A fully connected Multi-Layer Perceptron (MLP) with a decaying layer structure (256 $\to$ 128 $\to$ 64 $\to$ 32).
- Convolutional Neural Network (CNN): A 1D-CNN with three convolutional blocks using SiLU activation and Batch Normalization.
- Long Short-Term Memory (LSTM): A stacked LSTM network with two layers followed by an MLP head.
Hybrid Quantum Models:
- Quantum Neural Network (QNN): A variational circuit using data re-uploading (AngleEmbedding with Y-rotations) and a trainable StronglyEntanglingLayers ansatz. The quantum output is concatenated with PCA features and fed into a classical MLP.
- Quantum Convolutional Network (QCNN): Stacks parameterized "convolution" (entangling rotations) and "pooling" (qubit reduction) blocks. Features are extracted via expectation values and correlators, then processed by a classical MLP.
- Quantum Support Vector Machine (QSVM): Uses a fixed quantum feature map to compute a kernel (state overlap) which is then fed into a classical SVM solver. No end-to-end training of the quantum circuit occurs.
- Quantum LSTM (QLSTM): A hybrid approach where a variational circuit computes the gate pre-activations for the LSTM, while the recurrence logic remains classical.

Key Results
Performance is evaluated using Accuracy, ROC-AUC, F1-score, Precision, and Recall.

Classical Performance: The Artificial Neural Network (ANN) is the strongest performer, achieving 93.53% accuracy and a 0.9819 ROC-AUC. The CNN and LSTM follow closely (approx. 90% accuracy, 0.96 ROC-AUC). The classical SVM lags significantly (77.38% accuracy, 0.8186 ROC-AUC).
Quantum Performance: The Quantum Convolutional Network (QCNN) leads the quantum models with 90.89% accuracy and 0.9731 ROC-AUC. The Quantum Neural Network (QNN) is a close second (90.58% accuracy, 0.9706 ROC-AUC).
Underperforming Quantum Models: The QLSTM (74.07% accuracy) and QSVM (67.86% accuracy) perform markedly worse than their trainable hybrid counterparts.
Comparison: The best quantum models (QCNN/QNN) approach the performance of the best classical models (ANN) but do not surpass them. The trainable hybrid embeddings (QNN, QCNN) significantly outperform the fixed-kernel (QSVM) and recurrent quantum (QLSTM) approaches within the 8-qubit budget.

Significance and Claims
The paper explicitly frames this work as a controlled reference point rather than a claim of "quantum advantage."

Modest Claims: The authors state that the results demonstrate parity rather than superiority for quantum models in this specific regime. The study aims to provide explicit benchmark numbers for future comparisons.
Key Insight: Within the constraints of 8 qubits and PCA compression, trainable hybrid embeddings (QNN, QCNN) are more effective than fixed quantum kernels or recurrent quantum structures for this task.
Limitations Acknowledged: The authors highlight several constraints that prevent the results from being read as evidence of nontrivial pattern discovery:
1. Label Definition: Since the label is an analytic function of the inputs (invariant mass), the models are largely reconstructing a known function, inflating performance metrics.
2. Input Mismatch: Classical models use full features while quantum models use PCA-reduced features, mixing paradigm effects with representation effects.
3. Statistical Rigor: Results are based on a single split and seed, lacking error bars or cross-validation.
4. Reproducibility: The dataset description is brief, and code/seeds are not fully detailed in the text.

Conclusion
The study concludes that while hybrid quantum models can achieve high performance (near 91% accuracy) comparable to classical deep learning for trigger-like tasks, they do not yet offer a clear advantage. The value of the work lies in establishing a baseline for future research that must address rate-constrained operating points, latency measurements, and more rigorous statistical validation.

Technical Summary: Classical and Hybrid Quantum Machine Learning for Trigger-Like Event Selection

More like this