Hardware-Aware Quantum Kernel Design Based on Graph Neural Networks

Abstract: Quantum kernel methods have emerged as a promising direction in quantum machine learning (QML), offering a principled way to map classical data into high-dimensional quantum Hilbert spaces. While conceptually powerful, designing effective quantum kernels that adapt to both the target task and the constraints of near-term quantum hardware remains a nontrivial challenge. In this work, we propose HaQGNN, a hardware-aware quantum kernel design framework that integrates quantum device topology, noise characteristics, and graph neural networks (GNNs) to evaluate and select task-relevant quantum circuits. By predicting surrogate metrics related to fidelity and kernel performance, HaQGNN enables efficient circuit screening at scale. Feature selection is further incorporated to improve compatibility with limited-qubit systems and mitigate kernel degradation. Extensive experiments on three benchmark datasets, Credit Card, MNIST-5, and FMNIST-4, demonstrate that HaQGNN outperforms existing quantum kernel baselines in terms of classification accuracy. Our work highlights the potential of learning-based and hardware-aware strategies for advancing practical quantum kernel design on NISQ devices.