QAGO: Evolving Quantumness through Genetic Optimization of Quantum Circuits

Abstract Designing expressive yet hardware-efficient quantum circuits remains a central challenge in quantum machine learning (QML). Existing approaches to quantum circuit synthesis primarily optimize for classical performance metrics, often overlooking intrinsic quantum characteristics that govern circuit quality and generalization. This work introduces a Quantumness-Aware Genetic Optimization (QAGO) framework for circuit synthesis that integrates quantum properties such as entanglement and non-Clifford gate utilization into the evolutionary design process. We formulate circuit synthesis as both a single-objective optimization (SOO) and a multi-objective optimization (MOO) problem within a genetic framework. The SOO formulation explores quantumness-aware objectives balancing predictive performance and circuit-level quantum structure, while the MOO formulation approximates Pareto-efficient trade-offs between these goals. This dual perspective enables systematic analysis of how increasing quantum expressivity influences predictive behavior and resource cost. Experimental evaluations across three quantum kernel-based algorithms include Quantum Support Vector Machines (QSVM), Quantum Kernel Trainers (QKT), and Projected Quantum Kernels (PQK) demonstrate that circuits evolved under quantumness-aware objectives consistently outperform baseline kernel circuits such as ZZFeatureMap. Comprehensive benchmarking against classical baselines shows that QAGO achieves up to 2.5% AUC improvement on the Higgs Boson dataset and 1.5%–9.2% gains across six of nine targets in the LIT-PCBA drug discovery benchmark. Compared to prior work [1, 2] QAGO yields up to 4% improvement on Higgs and 5–20% gains across eight of nine targets. These results provide quantitative evidence that optimizing for quantumness enhances both performance and quantum expressivity of QML models, establishing QAGO as a principled framework for quantumness-aware circuit evolution.