Benchmarking machine learning approaches for polarization mapping in ferroelectrics using 4D-STEM.

Area of Science:

• Materials Science
• Electron Microscopy
• Machine Learning

Background:

• Four-dimensional scanning transmission electron microscopy (4D-STEM) offers atomic-scale material insights.
• Extracting physical properties like ferroelectric polarization from 4D-STEM data is challenging.
• Accurate polarization data is crucial for understanding ferroelectric functional properties.

Purpose of the Study:

• To benchmark machine learning models for automated polarization direction detection in ferroelectrics.
• To investigate the domain gap between simulated and experimental 4D-STEM data.
• To explore the potential of machine learning for identifying crystal structure defects.

Main Methods:

• Benchmarking ResNet, VGG, custom CNN, and PCA-kNN models.
• Training and testing models on 4D-STEM diffraction patterns from ferroelectric potassium sodium niobate.
• Utilizing data augmentation, filtering, and a prototype representation training regime.

Main Results:

• Models trained on synthetic data show high accuracy on synthetic patterns but struggle with experimental data.
• A custom training regime and PCA-based methods show promise in bridging the simulation-experiment domain gap.
• Periodic misclassification patterns suggest limitations in information content of some diffraction patterns.
• Model prediction irregularities correlate with crystal structure defects, indicating potential for defect detection.

Conclusions:

• Automating polarization detection in ferroelectrics using 4D-STEM is feasible but faces domain gap challenges.
• Advanced training strategies and data processing are crucial for real-world applicability.
• Supervised machine learning models show potential for identifying structural defects in materials.
• Further research is needed to validate the reliability of these methods for practical applications.