Deep PACBED: Multitask analysis of PACBED images using deep neural networks.

Area of Science:

• Materials Science
• Data Science
• Physics

Background:

• Scanning transmission electron microscopes (STEMs) generate large datasets of semiconductor images.
• Position-averaged convergent beam electron diffraction (PACBED) images are valuable for determining sample parameters.
• Manual measurement of these parameters is time-consuming and prone to inaccuracies.

Purpose of the Study:

• To investigate multitask deep neural networks for simultaneous analysis of multiple sample parameters from PACBED images.
• To compare the performance of multitask models against single-parameter models.
• To explore the impact of experimental data quantity on model performance.

Main Methods:

• Development and training of multitask deep neural networks.
• Utilizing synthetic PACBED images (III-V semiconductors, silicon) simulated via a multi-slice algorithm.
• Incorporating small amounts of experimental data for training and validation.
• Evaluating model performance based on mean absolute error for thickness, rotation, and mistilt, and accuracy for material classification.

Main Results:

• The best deep neural network achieved high accuracy: 4.19° for rotation (α), 0.43 milliradians for mistilt (β), and 2.93 nanometers for thickness (t).
• These error margins are comparable to experimental measurement uncertainties.
• Multitask models outperformed single-parameter models.
• Material classification was performed without errors on trained materials.

Conclusions:

• Multitask deep learning offers a rapid, consistent, and accurate method for analyzing large PACBED datasets from STEM.
• This approach significantly enhances the efficiency and reliability of semiconductor sample characterization.
• The findings suggest a promising direction for automated analysis in electron microscopy.