Evaluating Quantum Kernel Methods for Track-Based Classification in High-Energy Physics

This study investigates the feasibility and performance of a Quantum Support Vector Classifier (QSVC) for particle-track classification in the CLAS12 drift-chamber detector. The researchers aim to determine if quantum kernel methods can effectively capture complex, nonlinear correlations in detector-derived features that are difficult for classical models to resolve. The team encoded six-dimensional, intensity-weighted centroid features into six-qubit quantum states using a fully entangling ZZFeatureMap. They evaluated the model using both ideal statevector simulation and a calibrated hardware noise model (FakeMumbaiV2) to quantify the impact of decoherence on classification performance.

To handle the computational demands of high-energy physics datasets, the authors developed a multi-node, MPI-based HPC workflow. By decoupling quantum state preparation from kernel construction and caching statevector embeddings, they successfully scaled the evaluation to over 500,000 events. This approach enabled the construction of an exact kernel matrix for 100,488 training events, a scale more than an order of magnitude larger than previous quantum-kernel studies in this field. This methodology allows for rapid hyperparameter tuning and comparative analysis against classical baselines, including linear, polynomial, RBF, and sigmoid SVMs, as well as extremely randomized trees (ERT).

Under ideal simulation, the QSVC achieved a recall of 99.99%, identifying nearly all signal events with only 13 false negatives across 401,955 test events. While the ERT model achieved higher overall accuracy and background rejection, the QSVC’s near-perfect recall makes it uniquely suited for rare-event searches where signal loss is unacceptable. Under realistic hardware noise, the QSVC’s performance degraded, but it retained a recall of 99.51%. Geometric analysis revealed that hardware noise compresses the class-dependent kernel structure toward a maximally mixed state but preserves the relative ordering of similarities, explaining why the model’s discriminative power degrades gracefully rather than collapsing to chance.