Large Language Model-Driven Autonomous Quantum Circuit Design System

This article introduces an autonomous agent framework based on large language models for iterative quantum circuit design. The system forms a closed-loop workflow through seven components and has achieved excellent performance in quantum machine learning feature mapping and quantum chemistry Variational Quantum Eigensolver (VQE) tasks. The system integrates cutting-edge knowledge acquisition, code generation, critical evaluation, verification, storage, experimental assessment, and result review, providing a feasible paradigm for automated quantum circuit design.

Quantum circuit design has long relied on human experts' experience, with limited efficiency and constrained by the bottleneck of professional talents. As the number of qubits increases, the design space expands exponentially, making manual design difficult to handle. Large language models have strong code generation, reasoning, and knowledge integration capabilities, but their application in quantum circuit design needs to address challenges such as physical constraints, performance indicators, and iterative optimization. The research team proposes a framework that combines the generation capabilities of large language models with the closed-loop feedback mechanism of agents to achieve autonomous exploration and optimization.

The system consists of seven core components working collaboratively:

1. Exploration Component: Acquires cutting-edge research, open-source implementations, and technical documents from the network;
2. Generation Component: Converts requirements into executable quantum circuit code;
3. Discussion Component: Conducts theoretical analysis and comparison of designs based on literature;
4. Verification Component: Checks code correctness and executability;
5. Storage Component: Maintains design history and knowledge base, supporting experience reuse;
6. Evaluation Component: Obtains performance feedback through experiments;
7. Review Component: Analyzes experimental results and proposes improvement suggestions to guide iteration.

The system's effectiveness was validated in two major tasks:

• Quantum Feature Mapping: For quantum machine learning, the generated optimal feature mapping outperforms representative schemes in image classification benchmark tests, and its performance exceeds the classical radial basis function kernel method when extended to more qubits;
• Variational Quantum Eigensolver (VQE): The generated ansatz achieves accuracy comparable to chemically inspired and hardware-efficient schemes in seven molecular ground state energy estimation tasks, while meeting scaling constraints.

This work establishes the large language model-driven agent system as a feasible paradigm for automated quantum circuit design, demonstrating AI's ability to participate in iterative optimization workflows across scientific fields. Key design decisions include closed-loop feedback mechanisms, multi-component collaboration to simulate human design processes, and the introduction of domain knowledge constraints through literature-based criticism and verification. From a broader perspective, it provides new ideas for the automation of scientific discovery, where large language models can act as intelligent agents for active exploration of complex design spaces.

Future research directions include: expanding to more complex algorithm designs such as quantum error correction codes and quantum simulations; introducing multi-agent collaboration mechanisms; combining reinforcement learning to improve exploration efficiency; developing feedback signals that consider the noise characteristics of actual quantum hardware. This work marks a new stage of AI-assisted scientific discovery, and we look forward to AI playing a creative role in more scientific fields to accelerate technological innovation.