Atomic resolution convergent beam electron diffraction analysis using convolutional neural networks.

Area of Science:

• Materials Science
• Data Science
• Electron Microscopy

Background:

• Accurate sample thickness determination is crucial in scanning transmission electron microscopy (STEM) for quantitative analysis.
• Convergent beam electron diffraction (CBED) patterns provide thickness information but require sophisticated analysis.
• Traditional methods for thickness measurement can be time-consuming and complex.

Purpose of the Study:

• To develop and evaluate convolutional neural network (CNN) models for determining local sample thickness.
• To compare the performance of discrete classification and continuous regression CNN models using CBED patterns.
• To validate CNN-derived thickness measurements against quantitative high-angle annular dark-field (HAADF) STEM imaging.

Main Methods:

• Training two types of CNN models (discrete classification and continuous regression) on CBED patterns of SrTiO3.
• Generating training datasets from multislice simulations convolved with incoherent source broadening.
• Acquiring atomic resolution CBED patterns by balancing feature size, acquisition speed, and detector dynamic range.
• Utilizing computer vision techniques such as transfer learning and image augmentation.

Main Results:

• The regression CNN achieved 70% accuracy within 1 nm of HAADF thickness for samples < 35 nm, with a 1.0 nm root mean square error.
• The classification CNN, trained up to 100 nm, had 66% of results within a 2 nm increment of HAADF thickness.
• Both CNN models demonstrated effective thickness determination from CBED patterns.

Conclusions:

• CNN models offer a robust and efficient method for determining local sample thickness from CBED patterns in STEM.
• The developed approach provides a valuable tool for quantitative materials characterization at the nanoscale.
• This work highlights the potential of machine learning in advancing electron microscopy data analysis.