@article{Yang2026, 
author = {Yang Yang and Runsheng Zhao and Zhonghui Li and Cheng Xue and Zhaoyun Chen and Lei Qu},
title = {Review of quantum machine learning algorithms across diverse quantum computing architectures},
year = {2026},
journal = {Journal of National University of Defense Technology},
volume = {48},
number = {5},
pages = {1-19},
keywords = {quantum machine learning, quantum computing systems, noisy intermediate-scale quantum, parameterized quantum circuit},
url = {https://www.sciopen.com/article/10.11887/j.issn.1001-2486.25100007},
doi = {10.11887/j.issn.1001-2486.25100007},
abstract = {Significance The deceleration of Moore's Law has positioned quantum computing as a compelling alternative, leveraging superposition and entanglement for powerful high-dimensional computation. Although theoretical quantum advantages exist in tasks like boson sampling, the current Noisy Intermediate-Scale Quantum (NISQ) era limits high-fidelity algorithms such as Shor's and HHL due to noise and the absence of full fault-tolerance. Consequently, quantum machine learning (QML), especially its variational form that combines parameterized circuits with classical optimizers, has gained prominence. Its inherent noise resilience and strong generalization from limited data make it a leading candidate for achieving near-term practical quantum advantage, spurring research across quantum chemistry, drug discovery, finance, and artificial intelligence.  ProgressWhile QML broadly includes various quantum-classical hybrids, this review focuses on variational quantum machine learning. The standard workflow encodes classical data into quantum states, processes them through variational gates, and maps measurement outcomes to a classical objective function. A classical optimizer then adjusts circuit parameters iteratively, training the model. Evolving from early quantum-adapted algorithms like quantum support vector machines, the field has advanced toward sophisticated quantum neural networks (QNNs). Building on foundational work from the late 1990s, a rich ecosystem of quantum deep learning models, including Quantum Graph Neural Networks (QGNNs), Quantum Recurrent Neural Networks (QRNNs), and Quantum Generative Adversarial Networks (QGANs), has been developed and validated through classical simulations. However, the critical challenge lies in transitioning from simulation to execution on physical hardware. The quantum hardware landscape is notably diverse. Superconducting circuits use microwave pulses for scalable, high-precision control in cryogenic environments. Photonic systems manipulate photons with beam-splitters and phase-shifters, operating at room temperature with long coherence times. Trapped ions offer exceptional gate fidelities through laser control. Neutral atoms in optical tweezers provide flexible, scalable arrays. Solid-state approaches like color centers and quantum dots promise integration with existing technologies, while topological qubits aim for inherent error protection. This diversity in qubit design, gate operation, and system architecture creates a rich matrix of possibilities for implementing QML algorithms.  Conclusions and ProspectsThe central question is no longer solely algorithmic design but the synergistic co-design of algorithms with specific quantum hardware. The distinct properties of each platform directly influence QML algorithm performance, scalability, and application suitability. For example, the high gate fidelity of trapped ions may benefit deep QNNs, while the scalability of superconducting qubits could favor data-intensive tasks. The long coherence times of photonic or neutral atom systems may prove advantageous for specific hybrid algorithms. Thus, a systematic investigation is required to understand how platform-specific features—such as qubit connectivity, gate sets, coherence times, and error rates—shape QML algorithm characteristics across different application scenarios. This review addresses this core issue by providing a comparative analysis of QML algorithm progress across major quantum platforms. We examine physical qubit and gate implementations, analyze topological circuit structures, and synthesize research advancements in QML on these systems. By summarizing the respective advantages and ideal use-cases for QML on each platform, we can better chart future development trajectories for both quantum hardware and algorithms. This platform-centric perspective is essential for identifying the most promising pathways toward achieving tangible quantum advantages with machine learning, guiding future research toward optimal combinations of algorithms and hardware for specific practical applications.}
}