Graph-based Bayesian Optimization Discovers Variational Quantum Circuits for Cybersecurity Data Analysis

Designing effective quantum circuits remains a significant challenge in realising the potential of quantum computing for complex problems, and Prashant Kumar Choudhary from the Indian Institute of Technology (BHU), Nouhaila Innan and Muhammad Shafique from New York University Abu Dhabi, along with Rajeev Singh, present a new automated approach to tackle this issue. Their research introduces a system that intelligently discovers and improves quantum circuits using a technique called graph-based Bayesian optimisation, guided by a sophisticated ‘surrogate’ model built with graph neural networks. This method represents circuits as graphs, allowing the system to efficiently explore different designs and identify those best suited for classifying cybersecurity data, and the team demonstrates that their approach consistently produces circuits that are both simpler and more accurate than existing methods, even when subjected to realistic noise conditions. This achievement offers a scalable and interpretable pathway towards automated quantum circuit design, potentially accelerating progress in applying quantum computing to real-world challenges.

Automated Quantum Circuit Design via Neural Search This research pioneers an automated framework for discovering and refining variational quantum circuits, addressing a significant challenge in realising the potential of quantum computing for complex problems. Researchers represent circuits as graphs, enabling a structure-aware approach to optimisation, and employ a graph neural network (GNN) surrogate to predict circuit performance and uncertainty. This surrogate, specifically a lightweight GIN, utilises Monte Carlo dropout to calibrate epistemic uncertainty, improving the stability of ranking mutated candidate circuits during the search process. The methodology incorporates a tempered expected improvement acquisition function, integrating normalised complexity terms such as circuit depth, total gate count, and two-qubit gate counts. Crucially, the team implemented a transpiler-aware selection strategy, biasing the search toward circuits that maintain performance after mapping to realistic hardware. Experiments utilise a cybersecurity dataset, preprocessed with leakage-free techniques, to evaluate candidate circuits. The performance of the GNN-guided optimiser is benchmarked against other approaches, including MLP-based surrogates, random search, and greedy GNN selection. Robustness is assessed through comprehensive noise studies, including sweeps across thermal T1 and T2 relaxation times, analyses of amplitude and phase damping, and probability curves for depolarising and readout bit-flip noise. Detailed time breakdowns and Pareto frontiers for accuracy-complexity trade-offs are provided, and all artifacts, including code, configuration files, and exported circuits, will be made publicly available upon publication.,.

Graph Bayesian Optimisation of Quantum Circuits This research delivers a breakthrough in automated quantum circuit design, presenting a framework that discovers and refines variational circuits using graph-based Bayesian optimisation.

The team successfully represented circuits as graphs, enabling a graph neural network (GNN) surrogate to guide the search process, significantly improving upon traditional methods. Experiments demonstrate that this GNN-guided optimiser consistently identifies circuits with lower complexity while maintaining or exceeding the classification accuracy of benchmark approaches. The methodology centres on a pipeline that performs hardware-realistic circuit discovery end-to-end, leveraging a structure-aware GNN surrogate with calibrated uncertainty. This surrogate ingests circuit structure and costs, enabling principled interaction with compilers and hardware graphs. The framework utilises a cost-aware expected improvement acquisition function, incorporating realistic compilation settings using a SWAP-based bidirectional heuristic search algorithm. Robustness was assessed through a comprehensive noise study, evaluating performance across standard noise channels including amplitude damping, phase damping, thermal relaxation, depolarising, and readout bit-flip noise. The implementation is fully reproducible, with the team providing access to code, configuration files, and exported circuits, alongside surrogate metrics and thermal-sweep data.,. GNNs Discover Low-Complexity Quantum Circuits This research presents an automated framework for discovering variational quantum circuits, addressing a key challenge in applying quantum machine learning to complex, real-world data.

The team developed a system that uses graph-based Bayesian optimisation, guided by a graph neural network, to both discover and refine circuit designs. Circuits are represented as graphs, allowing the system to intelligently explore the design space and identify promising candidates. The method consistently finds circuits that are less complex and achieve classification accuracy comparable to, or exceeding, that of existing approaches. Evaluations using a cybersecurity dataset demonstrate the framework’s ability to identify circuits with favourable trade-offs between accuracy and complexity, specifically minimising the number of two-qubit gates and circuit depth. Furthermore, the system exhibits robustness to noise, maintaining performance even under conditions simulating realistic hardware limitations, and the inclusion of a decoherence-aware term during circuit selection further improves performance in noisy environments. The authors acknowledge that experiments primarily rely on simulations and suggest that future work should include evaluation on a wider range of backends and direct runs on quantum devices. Future research directions include incorporating multi-fidelity Bayesian optimisation, exploring richer sets of quantum operators, and pre-training structure-aware surrogates on larger datasets of circuit graphs. This work delivers a practical step toward scalable quantum machine learning by producing circuits that are accurate, efficient, and robust, and its modular design allows for adaptation to different datasets and hardware targets.,.

Automated Quantum Circuit Discovery For Classification This research presents an automated framework for discovering variational quantum circuits, addressing a key challenge in applying quantum machine learning to complex, real-world data.

The team developed a system that uses graph-based Bayesian optimisation, guided by a graph neural network, to both discover and refine circuit designs. Circuits are represented as graphs, allowing the system to intelligently explore the design space and identify promising candidates. The method consistently finds circuits that are less complex and achieve classification accuracy comparable to, or exceeding, that of existing approaches. Evaluations using a cybersecurity dataset demonstrate the framework’s ability to identify circuits with favourable trade-offs between accuracy and complexity, specifically minimising the number of two-qubit gates and circuit depth. Furthermore, the system exhibits robustness to noise, maintaining performance even under conditions simulating realistic hardware limitations, and the inclusion of a decoherence-aware term during circuit selection further improves performance in noisy environments. The authors acknowledge that experiments primarily rely on simulations and suggest that future work should include evaluation on a wider range of backends and direct runs on quantum devices. Future research directions include incorporating multi-fidelity Bayesian optimisation, exploring richer sets of quantum operators, and pre-training structure-aware surrogates on larger datasets of circuit graphs. This work delivers a practical step toward scalable quantum machine learning by producing circuits that are accurate, efficient, and robust, and its modular design allows for adaptation to different datasets and hardware targets. 👉 More information 🗞 Graph-Based Bayesian Optimization for Quantum Circuit Architecture Search with Uncertainty Calibrated Surrogates 🧠 ArXiv: https://arxiv.org/abs/2512.09586 Tags: