Abstract
One of the major objectives of experimental programs at the Large Hadron Collider (LHC) is the discovery of new physics, which requires identifying rare signals within immense backgrounds. As quantum technologies progress, Quantum Machine Learning (QML) could become a powerful tool for data analysis in high energy physics. In this study, using IBM gate-model quantum computing systems, we employ the quantum variational classifier (QVC) and the quantum kernel estimator (QKE) methods in two flagship LHC physics analyses: $t\bar{t}H$ (Higgs boson production in association with a top quark pair) and $H\rightarrow\mu\mu$ (Higgs boson decays to two muons). Using 10 qubits, we demonstrate that QML methods perform similarly to classical algorithms like Support Vector Machines (SVM) and Boosted Decision Trees (BDT) on simulators. Furthermore, tests on IBM quantum hardware show promising discrimination power comparable to simulator results, demonstrating QML's ability to differentiate signal from background in realistic physics datasets.
Introduction
High Energy Physics (HEP) is essentially a "needle in a haystack" problem. At the LHC, physicists search for incredibly rare events buried under a mountain of "background" noise from better-understood physical processes. While classical machine learning—specifically BDTs and SVMs—is currently the gold standard for this classification task, the sheer volume and complexity of future LHC data demand even more powerful tools. This study investigates whether the current generation of Noisy Intermediate-Scale Quantum (NISQ) computers is ready to handle realistic physics datasets and if they can offer a viable alternative to classical algorithms.
Methodology
The research employed two primary quantum approaches using 10 qubits:
- Quantum Variational Classifier (QVC): A hybrid algorithm where a quantum circuit encodes data into a quantum state, and a variational circuit (parameterized by "gate angles") is trained. A classical optimizer adjusts these angles to minimize classification error.
- Quantum Kernel Estimator (QKE): This method uses a quantum computer to map data into a high-dimensional Hilbert space to calculate a "kernel"—a measure of similarity between data points. This quantum-calculated kernel is then fed into a standard classical SVM.
To make high-dimensional LHC data compatible with 10-qubit systems, the researchers used Principal Component Analysis (PCA) to compress the kinematic variables of the particles into 10 essential training variables.
Datasets
The study utilized simulated data representing two "flagship" LHC analyses:
- $t\bar{t}H$: Higgs boson production in association with a top quark pair (focusing on the Higgs decaying to two photons).
- $H \rightarrow \mu^+\mu^-$: The rare decay of the Higgs boson into two muons.
Signal and background processes were generated using standard HEP tools: Madgraph5_aMC@NLO and Pythia6 for event generation, and Delphes for detector simulation. For hardware tests, datasets of 100 training events and 100 test events were used.
Key Results and Parity with Classical Benchmarks
The most significant finding is that QML is not merely a theoretical concept; it achieves performance nearly identical to highly optimized classical tools. The researchers used the AUC (Area Under the ROC Curve) to measure discrimination power, where 1.0 is perfect and 0.5 is random guessing.
As shown in the figure above, in the $H \rightarrow \mu^+\mu^-$ analysis, the QVC on the simulator reached an AUC of 0.83, slightly outperforming the classical BDT (0.80). For the $t\bar{t}H$ analysis, the quantum simulator achieved an AUC of 0.81, matching the high performance of classical benchmarks.
Hardware Resilience
One might expect real quantum hardware to underperform significantly due to noise (decoherence and gate errors). However, results from the ibmq_paris and ibmq_boeblingen processors were remarkably close to the noise-modeled simulators. This suggests that variational and kernel methods are robust in the NISQ era.
For the $H \rightarrow \mu^+\mu^-$ analysis, the QVC on real IBM hardware achieved an AUC of 0.81, nearly matching the 0.83 achieved on the simulator. Similarly, the QKE method on hardware reached an AUC of 0.85 for the $t\bar{t}H$ analysis, confirming that quantum kernels are highly effective for these datasets.
Limitations and Future Work
Despite these promising results, several constraints remain:
- Scale Constraints: Due to current hardware limits, the study was restricted to 10 qubits and small event samples. Classical models typically train on thousands or millions of events.
- No "Quantum Advantage" Yet: While performance is comparable, there is no speed-up in wall-clock time yet. Training and running these circuits on current hardware is significantly slower than classical execution.
Future work aims to scale these methods to more qubits and larger datasets as hardware improves. The ultimate goal is to find a "quantum advantage" where QML can either process data faster or identify signals that classical algorithms miss entirely.
Summary updated