Nonnegative matrix factorization incorporating domain specific constraints for four dimensional scanning transmission...

Area of Science:

• Materials Science
• Data Science
• Microscopy

Background:

• Modern electron microscopy generates large datasets requiring advanced analysis.
• Existing machine learning methods like PCA and NMF struggle with microscopy-specific data features.
• Domain-specific constraints are crucial for accurate interpretation of microscopy data.

Purpose of the Study:

• To develop a novel nonnegative matrix factorization (NMF) technique for analyzing four-dimensional (4D) scanning transmission electron microscopy (STEM) data.
• To integrate domain-specific constraints from electron microscopy into the NMF algorithm.
• To improve the decomposition and classification of material features in large microscopy datasets.

Main Methods:

• Proposed a constrained nonnegative matrix factorization (NMF) integrating spatial resolution and continuous intensity features.
• Applied the constrained NMF to simulated and experimental 4D STEM data.
• Optimized hierarchical clustering using diffraction similarity (polar coordinate transformation and uniaxial cross-correlation).

Main Results:

• Successfully decomposed 4D STEM data into interpretable diffractions and maps, surpassing PCA and standard NMF.
• Detected and classified nanometer-sized crystalline precipitates in ZrCuAl metallic glass using optimized clustering.
• Demonstrated superior performance over conventional machine learning techniques lacking domain knowledge.

Conclusions:

• The constrained NMF method effectively extracts material insights from large electron microscopy datasets.
• The developed technique accurately decomposes and classifies complex microstructures.
• This approach is broadly applicable to various characterization techniques and mitigates common machine learning artifacts.