AI system detects missing chip and restarts experiment on its own

A researcher pulled a chip out of the machine mid-experiment. Did not tell the system. Did not log it. Just reached in and took it. Qumus noticed the substrate was missing, figured out what had happened, and started over on a replacement chip without being asked.

That, according to a Princeton team, is the moment an AI quantum materials lab stopped being a demonstration and started being something you have to take seriously.

Qumus — described in a paper posted to arXiv May 18 — is what the team calls the first embodied AI quantum materials experimentalist. It runs inside a robotic mini-laboratory and handles the full experimental cycle: hypothesis generation, protocol planning, multi-step physical execution, real-time data analysis, and error recovery. No human in the loop for the core decisions. The paper's 17 authors, led by Sanfeng Wu at Princeton, say it has now done two things no AI has done before: created graphene from scratch, and fabricated a working atomically thin transistor by stacking two-dimensional materials without human hands touching the process.

The graphene flake took four hours and five optimization cycles. Qumus was given an empty experimental database and told to isolate a flake larger than 200 square micrometers — the size threshold that makes these materials useful for device fabrication. It explored a four-dimensional parameter space: stage temperature, contact time, massage cycles, and tape peeling speed. The result was a 245-square-micrometer flake. The target was arbitrary. The process was not.

The larger achievement is not the material itself. It is the closing of the loop between an AI that reasons about the physical world and a machine that acts on it. Large language models write papers. Robots exfoliate graphene. Getting them to work together as a single system — with error recovery — is what the field has been working toward.

Qumus runs a hierarchical multi-agent architecture under a central coordinator. There is a Project Manager that reads prior literature and proposes fabrication protocols. A Lab Manager that monitors inventory via computer vision. A Device Expert that designs layout. And a Processing Agent that executes. Below that, the system runs Atom Workflows — individual stage movements, camera adjustments, temperature changes — combined into Molecule Workflows, which combine into full Assembly Workflows for complete fabrication sequences.

The hardware is a compact modular workstation. Automated tape exfoliation for 2D crystal separation. Two robotic arms for moving materials between storage and temperature-controlled vacuum stages. An optical microscope with motorized focus for submicron alignment during layer transfer. Overhead cameras running YOLOv8 instance segmentation track tools and QR-coded carriers. A rule-based vision system estimates flake thickness from color distance in microscope images.

The chip-removal test is what makes the system worth writing about. During an hBN processing run, a researcher removed a silicon chip without notifying the system. Qumus detected the missing substrate through computer vision, generated a recovery strategy, and initiated re-exfoliation on a replacement chip. It also caught an instance where the language model misidentified an hBN flake as graphene — recognizing that no hBN had been recorded in the experimental database and adjusting its plan to isolate the correct material. These are not scripted error handlers. They are the system noticing something is wrong, reasoning about what happened, and fixing it.

The fabrication of the graphene field-effect transistor is the other concrete milestone. The Device Expert selected suitable graphene and hBN flakes from a substrate with pre-patterned metal contacts. The Processing Agent ran a 90-minute dry transfer procedure involving 30 physical operations and 18 decision stages. The system used real-time image analysis and Newton's rings detection to identify the contact point and assemble an aligned hBN-graphene heterostructure over the electrodes.

Mengdi Wang, a co-author at Princeton, put it plainly on LinkedIn: "This is not just an AI copilot for scientists." The distinction matters. A copilot assists. An experimentalist acts. Qumus was given a goal and left to run. The chip was a stress test the researchers staged to find out whether the system could handle real lab conditions — the kind where humans do unpredictable things without telling the machine.

The paper is on arXiv. It has not been peer-reviewed. The hardware is slow — the authors note mechanical movement speed, optical focusing, and thermal equilibration as current constraints. The graphene transistor performance is not compared against human-fabricated devices. The result is from one group. Those are real caveats and the paper makes them explicitly.

What the paper does not hedge on is the framework's generality. The authors write that the underlying multi-agent architecture is highly scalable. Faster robotic systems and improved thermal control would address the hardware bottlenecks. Linking multiple robotic laboratories via shared digital networks is not speculative — it is the obvious next step. The demo videos at qumus.ai show the system executing these sequences in real time.

Van der Waals stacking — the technique Qumus used to assemble the transistor — is how researchers build the heterostructures that go into quantum computing hardware, advanced sensors, and optoelectronic devices. It is finicky, hands-on work that requires a trained operator and long setup times. Automating it does not just speed up the process. It changes the geometry of a materials laboratory: the bottleneck shifts from synthesis to characterization. A system that can fabricate 24/7 will generate more samples than a conventional lab can analyze. The next constraint is not making the material. It is measuring what you made.

This is the direction the field is going. Human-in-the-loop materials science — where a researcher designs the experiment, executes the fabrication, takes the measurement, and draws the conclusion — has a ceiling. The trial-and-error iteration cycle is limited by how fast a person can work. Removing the human from the execution loop, even partially, opens a different operational mode: systems that run continuously, self-correct, and iterate faster than any individual researcher could manage. Qumus is a prototype. It is also a proof of concept for every autonomous lab that comes after it.

The paper: Shi, L. et al. "Qumus: Realization of an Embodied AI Quantum Material Experimentalist." arXiv:2605.18407, May 18, 2026.