An artificial intelligence system spent more than 25 hours moving molecules one by one to create a structure with electronic properties chosen in advance.
The experiment was conducted at Oak Ridge National Laboratory in the United States. To manipulate matter at such a small scale, the researchers use a scanning tunneling microscope. Its extremely fine tip can be moved above a surface and used to push molecules. This operation usually requires a great deal of time and precision from a human operator.
The researchers entrusted much of this work to an automated system. First, an artificial intelligence identifies carbon monoxide molecules deposited on a copper surface. The software then determines which ones need to be moved and to which locations. The microscope tip then performs the necessary movements.
The system also learns how to push each molecule. In particular, it adjusts certain electrical parameters and the speed of the movement based on the results of its previous attempts. A successful manipulation earns it a reward. This method, known as reinforcement learning, allows it to gradually improve its choices without a human controlling every movement.
To test the system as a whole, the team asked it to build a structure made up of 37 molecules arranged in a honeycomb pattern. This pattern artificially reproduces certain characteristics of graphene, a material made of a single layer of carbon atoms. The construction required about 900 operations over more than 25 hours.
Creating the shape was not enough. The researchers then measured how electrons behaved within it. They found an electronic signature characteristic of graphene, known as the Dirac point. In other words, the assembly of molecules did not merely have the right appearance: it also produced the desired electronic behavior.
The automation is not yet complete, however. The microscope tip can become unstable and sometimes needs to be repaired by an operator. Each movement also takes about one minute. The technique therefore remains a long way from an industrial process capable of rapidly manufacturing entire components.
The real interest lies in the possibility of testing new structures with far fewer human interventions. By positioning molecules precisely, researchers can modify how electrons move through the assembly. They can thus study artificial materials whose properties are defined by their organization at a very small scale.
A possible next step would be to completely reverse the process. Instead of providing a structure to build, researchers would specify only the desired electronic property. The system could then determine the corresponding molecular organization on its own and attempt to fabricate it with the microscope.