STEMDiff: A Wavelet-Enhanced Diffusion Model for Physics-Informed STEM Image Generation.

Area of Science:

• Materials Science
• Computational Science

Background:

• Machine learning (ML) application in STEM image analysis is limited by scarce annotated data.
• Deep generative models struggle to replicate high-frequency details in experimental STEM images.

Purpose of the Study:

• Introduce STEMDiff, a conditional diffusion model for generating realistic STEM images.
• Address the data bottleneck in STEM image analysis for ML applications.

Main Methods:

• Employed a conditional diffusion model (STEMDiff) with physical information embedding.
• Developed a Discrete Wavelet Transform-based skip-connection architecture to preserve high-frequency components and noise characteristics.
• Utilized binary labels from crystal structures to generate synthetic STEM images.

Main Results:

• Generated STEM images quantitatively nearly indistinguishable from experimental data, showing a 17-fold improvement over prior methods.
• Synthetic images retained ground truth structural information and experimental noise characteristics.
• Fully convolutional networks trained on synthetic data achieved high-precision atomic column detection in experimental WSe2 and graphene STEM images.

Conclusions:

• STEMDiff effectively overcomes the data scarcity issue in STEM image analysis.
• The model enables accurate atomic column detection without manual annotation, applicable to noisy experimental data.
• The approach is extendable to other scientific imaging modalities, potentially accelerating materials design for applications like water treatment.