37 molecules moved one by one: at Oak Ridge, the AI controls the tip of a microscope alone

A vision model and a reinforcement agent pushed molecules on copper to form a honeycomb network, in 900 iterations and 25 hours, without breaking the tool.

At the Centre for Nanophase Materials Sciences at Oak Ridge National Laboratory, a team led by P. Ganesh, Arthur P. Baddorf, An-Ping Li and Rama Vasudevan entrusted an AI with a task that physicists were carrying out so far by hand: moving individual molecules to the surface of a copper crystal to construct a structure chosen in advance.

The device rests on two bricks. A computer vision model of the YOLO family serves as eyes to the system: it identifies molecules in the images of the microscope. A reinforcement learning agent plays the role of brain and chooses the manipulations to try, learning not to damage the tip — the most fragile and costly part of the instrument.

Result published in ACS Nano : an assembly of 37 molecules arranged in perfect honeycomb, obtained after about 900 iterations spread over 25 hours, each movement taking about a minute. The stakes go beyond the demonstration: manufacturing matter with atomic precision interests fast electronics, advanced memories and, in the longer term, topological qubits.