[Paper Review] Agents for self-driving laboratories applied to quantum computing

Shuxiang Cao, Zijian Zhang+6 more|arXiv (Cornell University)|Dec 10, 2024

Scientific Computing and Data Management6 citations

TL;DR

The paper introduces k-agents, a knowledge-based multi-agent framework that uses LLM-based agents to automate laboratory experiments, demonstrated by calibrating and operating a superconducting quantum processor and generating entangled states with closed-loop feedback.

ABSTRACT

Fully automated self-driving laboratories are promising to enable high-throughput and large-scale scientific discovery by reducing repetitive labour. However, effective automation requires deep integration of laboratory knowledge, which is often unstructured, multimodal, and difficult to incorporate into current AI systems. This paper introduces the k-agents framework, designed to support experimentalists in organizing laboratory knowledge and automating experiments with agents. Our framework employs large language model-based agents to encapsulate laboratory knowledge including available laboratory operations and methods for analyzing experiment results. To automate experiments, we introduce execution agents that break multi-step experimental procedures into agent-based state machines, interact with other agents to execute each step and analyze the experiment results. The analyzed results are then utilized to drive state transitions, enabling closed-loop feedback control. To demonstrate its capabilities, we applied the agents to calibrate and operate a superconducting quantum processor, where they autonomously planned and executed experiments for hours, successfully producing and characterizing entangled quantum states at the level achieved by human scientists. Our knowledge-based agent system opens up new possibilities for managing laboratory knowledge and accelerating scientific discovery.

Motivation & Objective

• Motivate and address the challenges of automating laboratory experiments with multimodal and dynamic knowledge.
• Propose the k-agents framework to encode laboratory knowledge into agent-based systems without fine-tuning.
• Demonstrate autonomous calibration and operation of superconducting quantum processors using agent-driven workflows.
• Showcase how knowledge, translation, inspection, and execution agents enable closed-loop experiment planning and execution.

Proposed method

• Define knowledge agents to transfer laboratory knowledge via natural language.
• Implement code translation and procedure translation agents to convert natural language instructions into executable code.
• Introduce inspection agents that analyze experiment results and generate natural language reports.
• Develop an execution agent that decomposes procedures into an agent-based state machine and coordinates multi-step experiments.
• Benchmark translation accuracy against standard RAG methods and evaluate visual inspection efficacy with example figures.
• Apply the framework to automate single- and two-qubit gate calibration on a 16-qubit superconducting processor and to generate a GHZ state with tomography.

Experimental results

Research questions

• RQ1 How can a knowledge-based multi-agent system automate long-horizon, multimodal laboratory workflows without extensive fine-tuning?
• RQ2 Can LLM-based agents effectively translate natural-language experimental instructions into executable laboratory actions?
• RQ3 Does a closed-loop, agent-based state machine enable scalable, autonomous calibration and operation of superconducting quantum processors?
• RQ4 What is the impact of agent-based inspection (especially with visual examples) on deciding experiment success?
• RQ5 To what extent can such a framework produce entangled states and quantify fidelity comparable to human experts?

Key findings

• GPT-4o achieved the highest translation accuracy among tested models, with an accuracy of 97%.
• The system enabled fully automated recalibration of single-qubit parameters and automated discovery of two-qubit SIzzle gate parameters on a 16-qubit processor.
• Inspection agents using example figures yielded best performance in determining experiment success during visual inspection tasks.
• The workflow ran for up to 3 hours, exploring up to 100 experiments and testing up to 20 driving frequencies to identify optimal SIzzle gate parameters.
• Human-instructed GHZ state tomography yielded a fidelity of 83.83% for the density matrix.
• Randomized benchmarking on a validated single qubit achieved a gate fidelity of 99.869% ± 0.013%.