Manual identification and labelling of defects in scanning tunnelling microscopy (STM) images is a labour-intensive task requiring significant human effort, according to a recent paper on arXiv. To reduce this burden, researchers led by Kolev, Nikola L, Trouton, Max, Canova, Filippo Federici, Thornton, Geoff, Gao, David Z, Curson, Neil J, and Stock, Taylor J Z have proposed an automated approach that combines few-shot learning and unsupervised learning for segmentation of STM images.
The Challenge of Labelled Data Scarcity
Previous supervised methods for STM image analysis required large manually annotated datasets, which are difficult to obtain for new or unseen surfaces. The proposed technique removes this requirement, offering greater flexibility while still maintaining high accuracy. The paper notes that the approach is easier to adapt to an unseen surface compared to existing methods.
Technical Approach and Results
The researchers demonstrated the effectiveness of their model by using it to recognise atomic features on three distinct surfaces: Si(001), Ge(001), and TiO₂(110), including adsorbed AsH₃ molecules on the silicon and germanium surfaces. The model exhibits strong generalisation capabilities, and following initial training, can be adapted to unseen surfaces with as few as one additional labelled data point.
| Surface | Features Recognized |
|---|---|
| Si(001) | Adsorbed AsH₃ molecules |
| Ge(001) | Adsorbed AsH₃ molecules |
| TiO₂(110) | Atomic features |
The combination of few-shot learning (training on a small number of labelled examples) and unsupervised learning (extracting patterns from unlabelled data) enables this flexibility. The authors reported that the technique offers a significant improvement over fully supervised methods that require large datasets.
Implications for Materials Science
The work represents a significant step towards efficient and material-agnostic, automatic segmentation of STM images, according to the paper. By automating a labour-intensive task, the approach could accelerate research in materials science, enabling high-throughput analysis of atomic-scale defects without the bottleneck of manual annotation. This method is directly applicable to understanding physical and chemical processes at the level of single atoms and molecules, a core goal of STM imaging.
For enterprise technology leaders, the underlying principles — leveraging few-shot and unsupervised learning to overcome data scarcity — have potential applications beyond microscopy. Any domain where labelled image data is expensive or time-consuming to produce could benefit from similar techniques, though the paper focuses exclusively on STM. The demonstration of generalisation with minimal data points is a concrete metric that validates the approach's efficiency.