Unsupervised machine learning combined with 4D scanning transmission electron microscopy for bimodal nanostructural...

Area of Science:

• Materials Science
• Data Science
• Nanotechnology

Background:

• Scanning transmission electron microscopy (STEM) provides high spatial resolution for material analysis.
• Analyzing large datasets from techniques like four-dimensional STEM (4D-STEM) presents computational challenges.
• Unsupervised machine learning offers powerful tools for extracting meaningful information from complex data.

Purpose of the Study:

• To combine unsupervised machine learning with 4D-STEM for advanced crystal structure analysis.
• To develop optimized machine learning methods for processing large-scale 4D-STEM data.
• To apply these integrated techniques for characterizing nanostructures in materials.

Main Methods:

• Utilized unsupervised machine learning techniques, including non-negative matrix factorization (NMF) and hierarchical clustering.
• Developed custom software for data preprocessing, NMF, and hierarchical clustering tailored for electron diffraction and STEM imaging.
• Implemented hierarchical clustering with cross-correlation for rotation- and shift-corrected analysis.

Main Results:

• Successfully processed large-scale 4D-STEM data using dimensionality reduction techniques.
• Identified nanoscale crystalline precipitates (approx. 7 nm) within an amorphous matrix of Zr-Cu-Al metallic glass.
• Achieved detection of precipitates that were difficult to resolve with conventional STEM methods.

Conclusions:

• The integration of 4D-STEM with optimized unsupervised machine learning enables comprehensive bimodal analysis of material nanostructures.
• This approach enhances the capability for detailed crystal structure analysis at the nanometer scale.
• The developed methodology provides a powerful tool for materials characterization, particularly for complex systems like metallic glasses.