Machine Learning Puts Atom-Scale Defect Fabrication in 2D Materials on Autopilot

Key Findings · At a Glance

• ▸A fully autonomous loop fabricated metallic MoS-nanowire edge structures in monolayer MoS2 using a focused electron beam and machine learning feedback.
• ▸Directional nanowire growth succeeded in 65% of 42 experimental runs, comparable to 64% for earlier, less controllable automated seeding.
• ▸The atom classifier was trained on just 55 hand-labelled sites, a task the authors estimate takes an expert at most several hours.
• ▸Machine learning labelling showed a dramatically lower error rate than the group's earlier heuristic, rule-based framework.
• ▸Image decoding takes about 1 s on a GPU workstation and 3–5 s on an ordinary laptop.
• ▸Targeting sulfur rather than molybdenum atoms gave far more controllable vacancy formation.
• ▸The authors describe the framework as material-agnostic and extendable to other 2D materials and heterostructures.

Opportunity

Materials with tailored quantum properties can in principle be assembled atom by atom. In an aberration-corrected scanning transmission electron microscope (STEM), the same focused electron beam that images a lattice can knock individual atoms out of it. The paper describes this level of atomic precision as critical to advances in quantum sensing and communications. What has been missing is a process that runs reliably without a skilled operator deciding, frame by frame, where the beam goes next.

The team, led from the Center for Nanophase Materials Sciences at Oak Ridge National Laboratory, set out to close that loop. Their target structure is the MoS-nanowire (MoS-NW) edge: a metallic Mo6S6 nanowire attached to the semiconducting MoS2 lattice, a 1D-2D heterostructure with promising mechanical and electronic properties. The system acquires an image, decodes the identity and position of every atom, applies a predefined fabrication strategy to choose which sulfur sites to eject, and sends those coordinates to the scan coils.

Why Atom-Scale Fabrication Has Not Yet Delivered

Much of the earlier work on electron-beam atomic manipulation relied on manual control, with the user supplying both the strategy and the beam position. The group's own earlier automated method used intensity feedback and spiral scans, but it was restricted by the scan field of view and relied on favourable statistics for sulfur ejection, forming MoS-NW edges in 64% of attempts with little say over where they appeared. Automating the decision about where to place the beam requires reliable, rapid recognition of defects as the lattice deforms. The group's own heuristic classification criteria for such features were, in the authors' words, "tedious to implement, challenging to tune, and provided inconsistent results", while large computer vision models demand costly labelled data and transfer poorly from simulation to experiment.

Results

The decoding pipeline is deliberately small. An ensemble of U-net models, pretrained on simulated MoS2 images, locates atoms. A lightweight U-net trained on about 30 labelled images segments nanowire regions. A random forest, trained on 55 hand-labelled sites (21 Mo, 24 S and 10 sulfur-vacancy sites), sorts the remaining atoms into molybdenum, sulfur or vacancy-line classes by encoding each atom's distance and intensity relative to its five nearest neighbours. Against the group's older heuristic framework, the machine learning version produced a dramatically lower error rate and held up across varied image quality, beam conditions and carbon contamination. The authors cite this comparison as evidence of "the fragility of human-defined rules".

The Practical Applications

Assessment

The authors are candid about the limits. There is no temporal tracking of atoms between images; each frame is treated independently, so the system assumes the largest nanowire cluster remains the largest, and it sometimes wavers between small clusters early on. Targeting accuracy degrades during initial seeding as each ejected atom distorts the lattice. Failures also stem from uncorrected drift, defocus and stigmation shifts, and from contamination creeping into the field of view. Future work, they suggest, should add temporal tracking, smarter dose management to limit unwanted beam damage, simulation-informed training and AI agents that learn and adapt to changing microscope conditions.

Three analyst cautions follow. First, a 65% success rate is workable for research but some distance from a production process, and the paper itself describes it as comparable to the earlier method's 64%, which measured a different task. Second, the demonstration covers one material system and a handful of defect types in fields of view of 6–10 nm; the material-agnostic claim is an argument, not yet a result. Third, the paper does not measure the electronic or mechanical properties of the structures it builds, so their functional value rests on prior literature. The durable contribution is architectural: evidence that lightweight, quickly trained models are sufficient to close the loop between atomic imaging and atomic fabrication.