ORNL's AI steers a microscope tip to build a 37-molecule artificial graphene lattice

ORNL's AI makes the physical choices inside a running experiment, the ones a trained operator used to make by hand. According to the lab, building at this scale was laborious, and a single human error could break the delicate tools ^[13]. The system scores every attempted move. The reward is highest when a molecule reaches its intended site and much lower when it misses ^[3]. From those scores it settled on settings that push molecules into place without damaging the microscope tip ^[3].

The pace is slow. The lattice took about 900 iterations over 25 hours ^[6]. It works out to about 24 iterations for every molecule in the finished structure ^[8], or roughly 1.5 molecules placed per hour ^[9]. The trade-off is plain. Automation removes the operator's hours at the console, while each structure still occupies the instrument for about a day ^[6].

The sharpest objection comes from ORNL's own release. It says the AI worked more than 25 hours straight without a human operator ^[10]. Further down, it calls the process semi-automated, because a person must still occasionally step in to condition or repair the tip ^[11]. Both can be true if the AI ran the manipulation while people kept the tool working. The release does not say how often anyone intervened during the build. Ganesh Narasimha, who prepared the core automation software ^[7], chose more careful words. He said the work enables "atomically precise fabrication, with significantly reduced human input" ^[12]. The release also describes the system as building materials atom by atom ^[1], though the work it describes moves molecules across a copper surface ^[2].

We think the study supports a narrower version of the idea that automation is moving into the hands-on part of lab work. It holds for one instrument and one kind of task, and the output was verified. The team found a Dirac point in the lattice, evidence that the AI-built structure behaved like real graphene ^[4]. The AI also spelled out ORNL in individual molecules ^[5]. Both results come from the lab that built the system, describing its own paper in ACS Nano ^[1].

ORNL's longer-range claims concern what comes next. Its researchers point to "inverse design", where a scientist specifies an electronic property and the AI works out and builds the structure ^[14]. They call it a step toward topological qubits, the building blocks of ultra-stable quantum computers ^[14]. This study demonstrates the building half of that plan. "We expect this approach will accelerate realization and lead to new breakthroughs in our understanding of quantum states and future technology," Narasimha said ^[15]. For a research leader setting this quarter's budget, the usable result is narrower: a learned controller that chooses current, bias and speed for a fragile instrument and ran a build of about 900 iterations ^[3][6].