🔬 mesoscale physics

Autonomous Quantum Transport Measurements of 2D Semiconductors by an AI Agent

This paper presents an AI-driven workflow using the FermiLink agent harness to autonomously plan, execute, and analyze multi-day quantum transport measurements on 2D MoS2 devices, successfully determining spin-orbit coupling energies and mapping phase diagrams while ensuring instrumental safety.

Original authors: Brandon Bauer, Matthew Whalen, Kenji Watanabe, Takashi Taniguchi, John Q. Xiao, Tao E. Li, Wenjin Zhao

Published 2026-09-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Brandon Bauer, Matthew Whalen, Kenji Watanabe, Takashi Taniguchi, John Q. Xiao, Tao E. Li, Wenjin Zhao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quiet, freezing depths of a laboratory, scientists study materials that behave in ways ordinary matter does not. These are quantum materials, substances where electrons move in patterns so delicate they can only be observed at temperatures colder than deep space. To understand them, researchers must guide tiny electrical currents through microscopic devices while carefully adjusting magnetic fields and voltage. This process is slow, demanding, and requires a steady hand; a single mistake can ruin a fragile sample or damage expensive equipment. For decades, this work has relied entirely on human experts who spend days or weeks monitoring instruments, making split-second decisions, and interpreting complex data. The goal has always been to map out how these materials conduct electricity, revealing hidden properties that could one day power new technologies. However, the sheer complexity of these experiments means that progress is often limited by how fast a human can think and act.

A new approach is changing this landscape. Researchers have introduced an artificial intelligence agent capable of running these delicate experiments from start to finish without human intervention. This system does not just follow a rigid script; it plans, executes, and analyzes multi-day measurement campaigns on its own. The team tested this AI on devices made from molybdenum disulfide, a material that forms a single layer of atoms just a few nanometers thick. When these materials are stacked in different ways, their electrical properties change dramatically. In a single layer, the electrons behave in a specific way that splits their energy levels, while in two layers, the electrons can occupy different regions depending on the electric field applied. Understanding these behaviors requires gathering vast amounts of data across many variables, a task that is tedious for humans but ideal for an automated system.

The researchers built a framework called FermiLink to guide the AI. This system acts as a safety layer, ensuring that the AI never issues commands that could damage the delicate samples or the cryogenic instruments. The AI is given a simple goal, such as "measure the energy splitting in a single-layer device," and then it breaks this goal down into a series of logical steps. It decides what voltages to apply, how long to wait for the data, and how to interpret the results before moving to the next step. If a measurement fails or the data looks poor, the AI recognizes this and adjusts its plan, much like a human scientist would. This ability to adapt is crucial, as real-world experiments rarely go exactly as predicted. The system was tested on two types of devices: a single layer of molybdenum disulfide and a double layer.

In the first campaign, the AI focused on a single-layer device to determine the energy difference between two spin states of the electrons. Over the course of several days, the agent performed a sequence of measurements, starting by finding the right voltage range to turn the device on. It then applied magnetic fields to create patterns in the electrical resistance, which revealed the underlying quantum behavior. By analyzing these patterns, the AI calculated the energy splitting caused by the material's internal magnetic properties. It determined this value to be 13 millielectronvolts, a precise measurement that matched what human experts expected. The system also successfully identified a low-quality device early in the process, skipped unnecessary steps, and generated a report stating that no useful data could be obtained, demonstrating its ability to make judgment calls rather than blindly following instructions.

The second campaign was more complex, involving a two-layer device where the physics is even more intricate. Here, the AI had to map out how electrons distribute themselves between the top and bottom layers under different electric fields. The agent first measured the thickness of the insulating layers surrounding the device to calibrate its calculations. It then swept through a wide range of voltages and magnetic fields to create a detailed map of the material's behavior. By analyzing how the electrical resistance changed, the AI identified the boundaries where electrons switched from being shared between layers to being locked in one layer. It successfully constructed a phase diagram showing exactly where these transitions occurred, effectively charting a new territory in the material's behavior without human guidance.

The success of these experiments relies on three key design choices that keep the system reliable. First, the AI is trained to ask one specific question at a time, with clear rules for deciding when an answer is good enough to move forward. Second, the system uses a shared memory that allows different parts of the AI to remember what happened in previous steps, ensuring the entire multi-day campaign stays coherent. Third, the AI is separated from the physical hardware by a safety layer that prevents it from making dangerous mistakes, such as applying too much voltage. This setup allows the AI to handle the long, repetitive nature of these experiments while keeping the valuable equipment safe.

While this work represents a significant step forward, the researchers are clear about its current limits. The AI is not yet capable of inventing entirely new experimental strategies from scratch; it still relies on human experts to define the initial goals and provide the basic rules for how to measure specific materials. The system excels at executing and adapting these expert-defined workflows, but it cannot yet formulate its own scientific questions. Looking ahead, the team envisions a future where AI agents can build their own experimental plans by learning from past data and libraries of established methods. Until then, this autonomous agent serves as a powerful tool that can run complex, multi-day quantum transport measurements with a level of consistency and endurance that humans cannot match, opening the door to faster discovery in the world of quantum materials.

Technical Summary: Autonomous Quantum Transport Measurements of 2D Semiconductors by an AI Agent

Problem Statement
Quantum transport measurements are essential for discovering electronic phases in quantum materials, such as superconductors and topological insulators. However, acquiring and interpreting the high-dimensional datasets required for these studies—spanning multiple gate voltages, magnetic fields, and temperatures—is time-consuming and demands significant expertise. While Large Language Models (LLMs) and AI agents offer the potential to automate experimental tasks, current implementations face critical barriers. Specifically, quantum devices are delicate, and cryogenic instruments are fragile and valuable; unreliable agent actions can compromise both. Furthermore, existing literature largely reports single-step or human-in-the-loop measurements. End-to-end, multi-step quantum transport measurements conducted autonomously by AI agents without human intervention remain absent, particularly due to the instability of commercial agents over long-duration experiments (hours to days) and the lack of detailed decision trees for experimental workflows.

Methodology
The authors introduce a robust, agent-enabled framework for end-to-end quantum transport measurements, implemented through the FermiLink agent harness. The methodology relies on three core design principles:

The FermiLink Harness and Agent Loop: The system separates the AI agent's decision-making from direct hardware control. The harness manages the experimental loop:
- Planning: An AI agent (using models like Codex or Claude Code) designs detailed parameter sweeps based on a specific scientific question.
- Execution & Polling: The harness executes the measurement and polls the instrumental status during long waits, preventing the AI agent from timing out or becoming unstable.
- Analysis: Upon completion, data and project memory are passed to a new analysis agent for reasoning and post-processing.
- Iteration: This loop repeats, with each cycle answering a focused sub-question, until a comprehensive report is generated.
Hypothesis-Driven Agent Skills: The system utilizes a library of "Agent Skills"—templates containing domain knowledge, hard-coded control scripts, and post-processing logic derived from human expertise. These skills are iteratively refined through "dry runs" (simulated measurements on existing data) to ensure stability before being deployed on physical devices.
Hard Safety Constraints: To protect delicate 2D materials and expensive cryogenic equipment, hardware access is mediated by frozen Python scripts rather than the AI agents directly. Critical safety parameters (gate voltage limits, sweep rates, magnetic field ranges, and temperature limits) are hard-coded, prohibiting the agent from modifying them.

Experimental Campaigns and Results
The framework was validated using monolayer and bilayer $MoS_2$ devices, with autonomous campaigns lasting up to six days.

Monolayer $MoS_2$ Campaign:
- Objective: Determine the conduction-band spin-orbit coupling (SOC) energy and identify carrier density populations in $K$ and $K'$ valleys.
- Process: The agent autonomously executed a sequence of steps: determining the conducting gate-voltage window, performing Landau fan measurements to identify quantum oscillations, validating FFT frequencies via continuous magnetic field sweeps, and conducting Hall resistance measurements to determine band degeneracy.
- Outcome: The agent successfully determined that the lower conduction band is occupied below a carrier density of $n \approx 3.4 \times 10^{13} cm^{-2}$ and concluded an SOC splitting energy of 13 meV based on the given effective mass.
- Robustness Test: When applied to a low-quality device where FFT peaks were indistinct, the agent correctly identified the lack of clear data, skipped subsequent unnecessary steps, and generated a report concluding no peaks were accepted, demonstrating behavior akin to an expert experimentalist.
Bilayer $MoS_2$ Campaign:
- Objective: Map the layer- and valley-resolved phase diagram as a function of carrier density ( $n$ ) and electric field ( $E$ ).
- Process: The agent first calibrated the top/bottom hBN thickness ratio and absolute thicknesses using resistance and Hall measurements. It then mapped the $R_{xx}$ landscape across $n$ and $E$ at various magnetic fields to identify boundaries between layer-polarized and layer-shared regions. Finally, it performed Landau fan measurements at specific electric fields to assign valley occupations based on degeneracy rules (e.g., degeneracy of 2 for $K$ valleys, 12 for $Q$ valleys).
- Outcome: The agent successfully constructed a phase diagram distinguishing layer-polarized and layer-shared regions and assigned specific valley occupations to these regions.

Key Contributions

End-to-End Autonomy: The paper demonstrates the first instance of an AI agent performing multi-step, multi-day quantum transport measurements from initial instruction to final report without human intervention.
Safety-First Architecture: The FermiLink harness introduces a critical separation between the AI's "surface" (decision making) and the "hardware safety layer" (execution), ensuring that experimental failures do not damage equipment.
Iterative Skill Refinement: The methodology of using "dry runs" to iteratively update Agent Skills provides a pathway to stabilize AI agents for complex, long-duration scientific tasks.
Generalizable Framework: The system is designed to be adaptable to other types of experiments beyond $MoS_2$ , representing a step toward self-driving laboratories.

Significance and Claims
The authors claim that this work provides a valuable resource for the quantum materials community by demonstrating that AI agents can reliably execute complex, hypothesis-driven experimental workflows. The paper emphasizes that the framework converts open-ended experimentation into a sequence of verifiable decisions through the decision-tree structure of Agent Skills.

However, the authors remain modest regarding the current level of autonomy. They explicitly state that for new device types, human experts must still collaborate to develop robust Agent Skills using existing datasets. The current autonomy is defined as the execution and adaptation of expert-defined workflows, rather than the independent formulation of measurement strategies from first principles. The paper positions this work as a foundational step toward future self-driving laboratories where agents might eventually construct new skills, propose scientific questions, and search for novel phases independently.

More like this