Automating Experiments with Knowledge-Based AI Agents: A Case Study in Superconducting Quantum Processors

Experiment automation is crucial for scaling up quantum processors, as calibrating an increasing number of qubits becomes difficult to manage manually. Existing automation methods often depend on rigid, predefined procedures that lack flexibility, frequently requiring human intervention. Moreover, evaluating the success of experiments can be complex and difficult to describe purely in code. To address these difficulties, we propose an AI agent framework to automate these tasks. This framework employs multi-modal large language models to drive the execution of multi-step procedures based on natural language descriptions and oversee experiment results. We applied these methods to a superconducting quantum processor, demonstrating autonomous calibration of single-qubit gates, two-qubit gates, and the generation of GHZ states, achieving performance comparable to that of human scientists. This approach demonstrates the potential of AI agents to revolutionize experiment automation, opening a new era of routine practices in experimental physics.