Source-linked AI summary
Classical and Hybrid Quantum Machine Learning for Trigger-Like Event Selection on CMS Open Data: An Eight-Qubit, PCA-Constrained Benchmark
Tariq Mahmood, Muhammad Awais Rafique, Talab Hussain, Juan Pablo Perez Aguilar, Alfredo Raya, Muhammad Ahsan
TL;DR
High-energy physics triggers must retain rare signal events while rejecting overwhelming background under tight latency and bandwidth constraints. This paper benchmarks four classical and four hybrid quantum classifiers on a CMS open-data task using a shared protocol and an eight-qubit quantum budget. The artificial neural network is strongest overall, while the quantum convolutional network leads the quantum models; the study provides controlled reference numbers rather than evidence of quantum advantage.
Problem
Trigger-like event selection requires fast supervised discrimination between rare signal and overwhelming background under strict latency and bandwidth constraints.
Method
The study compares four classical and four hybrid quantum classifiers on CMS open data using invariant-mass and kinematic features under shared preprocessing, evaluation, and an eight-qubit, sixteen-component quantum budget.
Results
The artificial neural network is strongest at 93.53 percent accuracy and 0.9819 ROC-AUC, while the quantum convolutional network leads quantum models at 90.89 percent accuracy and 0.9731 ROC-AUC.
Takeaways & Limitations
Under this budget, trainable hybrid quantum embeddings outperform fixed-kernel and recurrent quantum variants, while the study shows parity rather than quantum advantage.
Takeaways & Limitations
The label is analytically derived from input components, and classical models receive the full feature set while quantum models receive sixteen principal components.
Abstract
from arXiv · showhide
Event triggering sits at the heart of high-energy physics, where the rare events of interest must be retained while an overwhelming background is discarded under tight latency and bandwidth budgets. This work compares four classical machine learning models, namely a support vector machine, an artificial neural network, a convolutional network and a long short-term memory network, with four hybrid quantum counterparts, on a trigger-like binary classification task built from CMS open data. The label is defined by an invariant-mass window, and the inputs combine reconstructed kinematics with physics-motivated derived variables: the pseudorapidity difference, the wrapped azimuthal difference, the angular separation and the total transverse momentum. The quantum models run under a fixed resource budget of eight qubits, a principal-component compression to sixteen features and state-vector simulation. Every model shares the same stratified split, the same preprocessing and a common decision threshold, and performance is reported through accuracy, ROC-AUC, F1-score, precision and recall. The strongest classical model is the artificial neural network, at 93.53 percent accuracy and 0.9819 ROC-AUC, while the strongest quantum model is the quantum convolutional network, at 90.89 percent accuracy and 0.9731 ROC-AUC, with the quantum neural network close behind. The quantum-kernel and recurrent quantum approaches trail both, which places the trainable hybrid embeddings ahead within this budget. The study is meant as a controlled reference point rather than a claim of quantum advantage.
1 Introduction
High-energy physics triggers must rapidly retain rare signal events while discarding overwhelming background under strict latency and bandwidth limits. This study compares classical and hybrid quantum models on the same CMS-based trigger-like classification task and resource-constrained protocol.
- Triggers inspect collisions online and retain only events worth recording under strict latency and readout-bandwidth limits.
- Supervised classification maps each event to a signal or background label, while quantum machine learning encodes data in circuits and measures a score.
- The study asks how classical and hybrid quantum models compare when given identical data, task, evaluation protocol, and a realistic quantum resource budget.
- The benchmark uses CMS open data, an invariant-mass-window label, reconstructed kinematics, derived variables, and eight models trained under a shared protocol.
- Quantum models are restricted to eight qubits, sixteen principal components, and state-vector simulation.
2 Related work
Prior work applies classical, quantum-kernel, hybrid, and model-independent methods to collider selection and classification. This study instead emphasizes a controlled comparison of classical and hybrid quantum families under a fixed modest quantum budget.
- Classical deep learning and quantum-inspired methods have been applied to online selection, event classification, collider simulation, and real-time triggering.
- Quantum-kernel studies report competitive background rejection or area under the curve across collider and B-meson classification tasks, with hardware results sometimes lower than simulator results.
- Model-independent approaches use variational autoencoders and foundation-model representations to identify anomalies or learn transferable collider features.
- The present study fixes the task, dataset, and protocol, then compares classical and hybrid quantum families under an explicit modest quantum budget.
3 Data and task definition
The task classifies CMS events using an invariant-mass window and kinematic features, including derived angular and transverse-momentum quantities. The label is analytically tied to input components, which limits how results should be interpreted.
- The dataset contains 31,892 CMS open-data events with 21 columns for a binary trigger-like classification task.
- The di-object invariant mass is computed from the measured four-vector components of two reconstructed objects.
- Events receive y = 1 when 3.0 ≤ M ≤ 3.2 GeV, the J/ψ resonance window, and y = 0 otherwise.
- Inputs combine raw kinematics with Δη, wrapped Δϕ, angular separation ΔR, and total transverse momentum pT,tot; missing-value events are removed globally.
- Because the label is an analytic function of input components, much achievable performance follows directly from the task definition rather than nontrivial pattern discovery.
- Models use an 80/20 stratified train-test split, validation held out from training, standardized features, shared seeds, and a fixed decision threshold of 0.5.
4 Methods
The study implements classical and hybrid quantum classifiers for trigger-like binary event selection, using shared preprocessing and a fixed eight-qubit quantum resource budget. Classical models include maximum-margin, fully connected, convolutional and recurrent architectures, while quantum models use variational circuits, quantum convolutions, state-overlap kernels and quantum-enhanced recurrence.
- Classical models: The support vector machine separates classes with a maximum-margin hyperplane and an RBF kernel over support vectors.The RBF kernel implicitly maps inputs to a higher-dimensional feature space; learned weights mostly vanish except for support-vector contributions.
- Classical models: The artificial neural network uses fully connected layers with nonlinear activations to map features to a trigger probability.Its implementation is a multilayer perceptron with dense hidden layers arranged as Input → 256 → 128 → 64 → 32 → Output.
- Classical models: The convolutional network reshapes tabular features into a one-dimensional sequence and applies shared-weight convolutional blocks before classification.The PyTorch implementation contains three Conv1d blocks with increasing channel counts, batch normalization and SiLU activations.
- Classical models: The LSTM controls information flow through input, forget and output gates while maintaining a recurrent memory state.The binary classifier uses two sequential LSTM layers and feeds the original feature vector as a sequence of length one.
- Quantum models: Every quantum model uses eight qubits, sixteen PCA-compressed features and state-vector simulation.Standardized inputs are compressed to sixteen components before encoding into the eight-qubit register.
- Quantum models: The quantum neural network encodes data, applies a trainable variational ansatz and measures an observable, whereas the QCNN alternates local convolution and pooling unitaries.The QNN is trained end-to-end with a statevector simulator; the QCNN reduces active qubits through pooling blocks before measurement.
- Quantum models: The quantum support vector machine replaces the classical kernel with an overlap between encoded quantum states and then uses a classical margin classifier.Its simulated statevectors define a fidelity kernel, whose Gram matrix is supplied to a precomputed-kernel SVM.
- Quantum models: The quantum LSTM computes gate pre-activations with a variational circuit while retaining the recurrence and memory update classically.Circuit outputs provide the candidate pre-activations, which enter the classical cell-state and hidden-state updates.
5 Results
The artificial neural network is the strongest classical model, while the quantum convolutional network leads the quantum models under the shared benchmark. Across both families, trainable neural embeddings outperform the weaker support-vector and recurrent approaches within the imposed resource budget.
- Classical models: 93.53 percent accuracy and 0.9819 ROC-AUC make the artificial neural network the strongest classical model.Its precision is 0.9104 and recall is 0.9458, indicating balanced performance at the chosen operating point.
- Classical models: The artificial neural network leads the classical ROC and F1 comparisons, with convolutional and recurrent networks near 0.96 area under the curve and the support vector machine trailing.The learning curves also show smooth validation behavior for the artificial neural network, consistent with stable convergence.
- Quantum models: 90.89 percent accuracy and 0.9731 ROC-AUC make the quantum convolutional network the strongest quantum model.The quantum neural network follows closely at 90.58 percent accuracy and 0.9706 ROC-AUC.
- Quantum models: Both leading quantum models reach recall above 0.95, whereas the quantum long short-term memory and quantum support vector networks are markedly weaker.The recurrent and kernel variants reach 74.07 percent and 67.86 percent accuracy, respectively.
- Quantum models: The quantum convolutional network shows high signal efficiency with controlled misclassification, while the recurrent and kernel variants accumulate many misclassified events.These confusion-matrix results indicate limited separability for the weaker quantum constructions under the imposed constraints.
6 Discussion
The benchmark favors trainable neural and hybrid quantum embeddings over fixed-kernel and recurrent alternatives, while several design limitations constrain interpretation. Trigger relevance is better captured by signal efficiency and background acceptance than by accuracy alone.
- Quantum models: 0.9731 ROC-AUC is achieved by the quantum convolutional network, ahead of QNN at 0.9706, QLSTM at 0.8325 and QSVM at 0.7340.The convolutional and quantum neural networks reach accuracy near 91 percent and recall above 0.95.
- Quantum models: The quantum support vector machine is weakest because its fixed, non-adaptive kernel requires quadratic kernel construction and training-event subsetting.A fixed kernel captures the decision boundary less effectively than a trainable hybrid under the stated constraints.
- Quantum models: The quantum long short-term memory network trails because principal components lack physical sequence structure and recurrent circuit evaluations complicate optimization.These constraints produce lower and less consistent area under the curve within the fixed epoch and simulation budgets.
- Limitations: The invariant-mass label is an analytic function of input components, so high performance partly follows directly from reconstructing a known function.A cleaner benchmark would withhold those components or use an independent criterion, alongside an explicit analytic mass-cut baseline.
- Limitations: The comparison mixes model paradigm with representation because classical models use the full feature set whereas quantum models use sixteen principal components.Resource-matched inputs would separate these effects.
- Limitations: Single-split, single-seed reporting without error bars or repeated runs leaves close quantum-model gaps potentially within statistical noise.Several seeds, uncertainty estimates and significance tests are needed for a reliable ranking.
- Trigger implications: Trigger evaluation should emphasize signal efficiency and background acceptance, with thresholds selected from rate and bandwidth constraints rather than fixed accuracy.High recall and ROC-AUC suggest threshold movement can expose a rate–efficiency trade-off, while deployment studies should measure latency and throughput.
7 Conclusion
Under a shared protocol, the artificial neural network is strongest overall, while trainable hybrid quantum embeddings lead the quantum models. The study supports parity rather than quantum advantage and serves as a controlled reference point for further deployment-oriented evaluation.
- 7 Conclusion: 93.53 percent accuracy and 0.9819 ROC-AUC make the artificial neural network strongest overall, while the classical support vector machine is weakest.The classical convolutional and recurrent networks are close behind the artificial neural network.
- 7 Conclusion: Roughly 90.6 to 90.9 percent accuracy and areas near 0.97 place the quantum convolutional and neural networks ahead of fixed-kernel and recurrent quantum variants.This ordering holds under the eight-qubit and sixteen-component budget.
- 7 Conclusion: The benchmark indicates parity rather than quantum advantage in this regime.Its value is the controlled reference numbers and the explicit identification of rate, efficiency, latency and robustness studies still needed.