Imagine a construction site where the bricks are individual molecules and the “cranes” are microscopic needles so sharp they can feel a single atom. For decades, building at this scale was a laborious and time-consuming task where a single human error could break the delicate tools. Now, researchers at the Department of Energy’s (DOE) Oak Ridge National Laboratory (ORNL) have handed the controls to an artificial intelligence that can “learn” how to build these materials autonomously, working more than 25 hours straight without a human operator.
In a landmark study published in ACS Nano, the ORNL researchers, from the Center for Nanophase Materials Sciences (CNMS), describe the creation of a fully automated system capable of building functional materials atom-by-atom. The effort grew from a conceptual framework shaped by P. Ganesh and Arthur P. Baddorf, together with An-Ping Li and Rama Vasudevan, who supervised the implementation and guided the interpretation of the results.
Turning that vision into a working platform required an unusual blend of software and hands-on experimental skill. Ganesh Narasimha prepared the core automation software based on inputs from Vasudevan, while Mykola Telychko developed the experimental setup; the two then jointly carried out the manipulation trials and assembled the resulting dataset. Wooin Yang helped design the AI’s object-detection workflow and performed the critical spectroscopic studies used to verify the materials. Narasimha said the collective effort enables “atomically precise fabrication, with significantly reduced human input.” He added that this effort is shifting the research paradigm from discovering materials in nature to engineering specific “artificial lattices,” which are orderly, repeating patterns of atoms or molecules that form a material’s underlying framework, with tailored electronic behaviors.
AI Vision and Control
The team developed AI models that served as microscopic “eyes” and a “brain.” As the “eyes,” a computer vision model called YOLO (“You Only Look Once”) was used to rapidly detect molecules on a copper surface. As the “brain,” reinforcement learning enabled the model to evaluate different strategies for moving molecules and receive a numerical reward based on how well each attempt succeeded. The reward was highest when a molecule reached the intended target site and much lower when it missed. By linking higher rewards to the actions that produced them, the model learned the best combination of electrical current, bias and manipulation speed to nudge molecules into place without damaging the microscope tip.
The researchers proved their designer results work by using the AI to build an artificial graphene lattice made of 37 molecules in a perfect honeycomb pattern. Most importantly, they confirmed the structure worked by finding a Dirac point, a unique electronic signature, proving the man-made material behaved exactly like real graphene. To demonstrate versatility, the AI even spelled out ORNL using individual molecules.
Science is entering an era in which matter itself is a programmable resource. By combining AI decision-making with atomic-scale precision, researchers can force electrons to behave in custom-designed ways that do not exist in the natural world. This capability, showcased in the ACS Nano study, enables the engineering of materials needed for high-speed electronics and advanced memory devices from the ground up.
The process is currently semi-automated because a human operator must still occasionally step in to condition or repair the microscope tip if it becomes unstable. Additionally, construction is time-sensitive; building the 37-molecule lattice required approximately 900 iterations over 25 hours, with each individual manipulation taking roughly one minute.
The researchers point to a future shift toward “inverse design,” where scientists define a desired electronic property and the AI autonomously determines and builds the required structure. This is a vital steppingstone toward building topological qubits, building blocks for ultra-stable quantum computers. “We expect this approach will accelerate realization and lead to new breakthroughs in our understanding of quantum states and future technology,” Narasimha said.
The experimental research was primarily supported by the DOE Office of Basic Energy Sciences, Scientific User Facilities Division, as part of the QIS Infrastructure Project “Precision Atomic Assembly for Quantum Information Science” (FWP ERKCZ62). The algorithmic development was supported by CNMS. Portions of the research used resources at the CNMS, which is a DOE Office of Science User Facility at ORNL. By combining state-of-the-art nanoscience tools, deep technical expertise and ORNL’s computing capabilities, CNMS helps make precision, AI-driven atom-by-atom assembly possible.
Source: