Quantum complete graph self-attention network for particle flow classification

Abstract Particle flow classification is essential in high-energy and nuclear physics. This work proposes the quantum complete graph self-attention network (QCGAT), which selects particles with high transverse momentum and uses a quantum self-attention mechanism for feature extraction. Unlike previous methods, the proposed QCGAT framework introduces a non-measurement quantum self-attention mechanism within an entangled quantum network, enabling the direct embedding of attention coefficients into quantum state amplitudes and thereby avoiding quantum-to-classical conversion. Furthermore, this work introduces a quantum distributed dimensionality reduction architecture that specifically enables tunable cross-scale data adaptation. Experiments on the Top Quark Tagging dataset show that QCGAT achieves 83.5% accuracy and 88.7% AUC, surpassing hybrid models by 1% and 0.4%. Compared with traditional quantum encoder baseline, it improves accuracy by 0.7%; when compared with traditional quantum encoder baselines having similar or fewer parameter counts, it achieves up to 1.6% higher accuracy. These results highlight the effectiveness and potential of QCGAT in quantum machine learning for particle physics.