A deep convolutional neural network to analyze position averaged convergent beam electron diffraction patterns.

Area of Science:

• Materials Science
• Computational Science
• Data Science

Background:

• Convergent beam electron diffraction (CBED) is crucial for materials characterization.
• Manual analysis of CBED patterns is time-consuming and labor-intensive.
• Automating CBED analysis is essential for handling large datasets, such as those from 4D STEM.

Purpose of the Study:

• To develop and validate deep convolutional neural networks (CNNs) for automated analysis of CBED patterns.
• To enable rapid and accurate measurement of sample thickness and tilt from diffraction data.
• To assess the performance and generalizability of the CNN approach.

Main Methods:

• A series of CNNs were designed to first calibrate CBED pattern parameters (zero-order disk size, center, rotation).
• Subsequent networks were trained to measure sample thickness and tilt using the aligned diffraction data.
• The methodology included exploring network response to various pattern features and varying experimental parameters (thickness, tilt, dose).

Main Results:

• The CNNs successfully calibrated CBED patterns without data preprocessing.
• Accurate measurements of sample thickness and tilt were achieved.
• The developed network processed patterns at approximately 0.1 s/pattern, orders of magnitude faster than brute-force methods.
• The approach demonstrated robustness across different materials and orientations, with potential for hybrid methods.

Conclusions:

• Deep convolutional neural networks offer a highly efficient and accurate method for automated CBED pattern analysis.
• This AI-driven approach is suitable for processing large-scale datasets, particularly from 4D STEM.
• The developed methodology provides a foundation for accelerating materials characterization and analysis.