A small-dataset-trained deep learning framework for identifying atoms on transmission electron microscopy images

Area of Science:

• Materials Science
• Microscopy
• Artificial Intelligence

Background:

• Transmission electron microscopy (TEM) is crucial for atomic-scale imaging.
• Analyzing noisy or low-dose TEM images for precise atom identification remains challenging.
• Current methods struggle with limited datasets and beam-sensitive materials.

Purpose of the Study:

• To develop a deep learning (DL) approach for accurate atom identification in noisy TEM images.
• To enable element discernment at the atomic level.
• To train DL models effectively with small experimental datasets.

Main Methods:

• A deep learning network was designed with a specialized loss function for feature extraction.
• The network was trained on a small dataset of approximately 30 experimental TEM images (256x256 pixels).
• Model performance was validated using graphene and PbTiO3/SrTiO3 multilayer structures, including simulated noisy images.

Main Results:

• The DL network achieved accurate atom identification and element discernment on noisy TEM images.
• The model demonstrated robustness by correctly analyzing images with randomly lost pixel intensities.
• Successful application to resolve structural defects in graphene and polar structures in multilayers was confirmed.

Conclusions:

• A DL framework trained on small datasets can effectively identify atoms in noisy TEM images.
• This approach offers potential for analyzing beam-sensitive materials and dynamic atomic processes.
• The DL methodology provides a valuable framework for diverse scientific applications requiring small-dataset training.