Researchers from Princeton University, University of Michigan, California State University and Japan's National Institute for Materials Science have introduced Qumus, an embodied AI system that can autonomously create graphene and fabricate atomically thin graphene devices in a robotic mini-laboratory.
Qumus AI architecture and fully robotic minilab. a Defining characteristics of an AI experimentalist. b Key self-evolving modules of Qumus, including LLM-agents, memory and knowledge systems, and skills including instrumental workflows and materials/devices realization recipes. c Qumus architecture for efficient multi-agent collaboration and robust performance. d A compact, fully robotic minilab consisting of vacuum- and temperature-controlled stages for 2D material mechanical exfoliation, optical flake search, flake transfer and stacking, along with robotic arms, storage modules, cameras, and microscope systems. Image from : arXiv
Qumus is built around the complete graphene workflow: from exfoliating bulk crystals to isolating single-layer flakes and stacking them into functional van der Waals (vdW) devices, all without human intervention. The system combines generative AI, computer vision and robotics to handle the labor-intensive steps that typically limit graphene research, such as flake discovery, thickness assessment and submicron alignment during transfer.
In a key benchmark, Qumus was tasked with isolating a graphene flake larger than 200 μm² starting from an empty experimental database, directly targeting a realistic device-ready flake size rather than an idealized test. Over a continuous run of more than 4 hours, it explored a four-dimensional parameter space - stage temperature, contact time, massage cycles and tape peeling speed - using iterative closed-loop optimization. After five optimization cycles, the system successfully produced a graphene flake with an area of 245 μm², satisfying the predefined size constraint.
Beyond flake production, Qumus demonstrates fully automated fabrication of a graphene field-effect transistor (FET) via vdW stacking, representing the first AI-driven construction of such an atomically thin nanodevice. The Device Expert Agent selects appropriate graphene and hexagonal boron nitride (hBN) flakes on a substrate with pre-patterned metal contacts, ensuring that the chosen flakes and geometry are compatible with the intended FET layout. The Processing Agent then executes a 90-minute dry-transfer sequence consisting of 30 physical operations and 18 AI-controlled decision points, guided by real-time image analysis and Newton’s rings detection to locate the correct contact point and achieve precise hBN-graphene alignment over the electrodes.
This experiment showcases not only automated flake manipulation but also end-to-end device assembly: Qumus moves from raw exfoliated material to a complete graphene FET within about 1.5 hours of robotic processing time. In parallel, the system logs detailed metadata for each step - exfoliation parameters, alignment conditions, optical signatures and device records - creating a structured digital trail that can be mined by machine learning models to optimize future graphene device fabrication.
Qumus is organized as a hierarchical multi-agent large language model (LLM) framework that mirrors a human graphene lab team. A central coordinator interprets user goals (for example, “isolate >200 μm² graphene” or “fabricate a graphene FET”), breaks them into subtasks and assigns them to specialized agents such as the Project Manager, Device Expert, Lab Manager and Processing Agent. The Processing Agent itself runs three workflow tiers: Atom Workflows (primitive actions like stage motion, focusing and temperature control), Molecule Workflows (compound tasks such as tape exfoliation and flake transfer) and Assembly Workflows that chain these into full graphene-device fabrication protocols.
On the hardware side, a tape exfoliation system deposits crystal layers onto silicon chips using automated Scotch tape processing, while two robotic arms shuttle materials between storage positions and temperature-controlled vacuum stages. Overhead cameras with YOLOv8 instance segmentation track tools and QR-coded carriers, and a rule-based microscope-vision pipeline analyzes RGB images, performs edge detection and estimates flake thickness via color-distance metrics to distinguish graphene from other 2D materials. These sensing and control loops allow Qumus to adapt in real time, for example when correcting a mislabelled hBN flake or recovering from the unannounced removal of a silicon chip mid-process by re-planning exfoliation on a new substrate.
By fully automating both the isolation of device-scale graphene flakes and the assembly of a working graphene FET, Qumus directly tackles the human-limited bottlenecks of 2D material device fabrication. Replacing manual flake hunting and alignment with systematic, data-logged workflows enables more reproducible graphene processes and faster screening of vdW heterostructures such as hBN–graphene stacks. The authors note that current throughput is constrained mainly by mechanical motion, optical focusing and thermal equilibration times, suggesting that faster robotics and optimized optics could further accelerate graphene device production on this platform.
Looking ahead, integrating the same architecture into inert-atmosphere gloveboxes would extend the approach from graphene and hBN to more air-sensitive 2D materials, broadening the range of vdW heterostructures and quantum devices that can be fabricated autonomously. Coupled with shared digital databases across multiple robotic labs, this kind of embodied AI could help move graphene and related materials from bespoke laboratory demonstrations toward scalable, data-driven device engineering.
