Dimensionality reduction and unsupervised clustering for EELS-SI.

Area of Science:

• Materials Science
• Data Science
• Spectroscopy

Background:

• Electron Energy Loss Spectroscopy Spectrum Imaging (EELS-SI) generates large, complex datasets.
• Differentiating subtle spectral features in EELS-SI is challenging.
• Existing methods may require prior knowledge or lack efficiency for large datasets.

Purpose of the Study:

• To develop a novel, data-driven machine learning strategy for differentiating distinct spectra within EELS-SI datasets.
• To enable efficient clustering and analysis of EEL spectra with similar fine structures.
• To facilitate the investigation of specific fine structures and their spatial distribution in materials.

Main Methods:

• Application of linear and nonlinear dimensionality reduction techniques to project EEL spectra into a low-dimensional space.
• Utilizing a density-based clustering algorithm for distinguishing meaningful data clusters.
• A novel combination of machine learning algorithms for spectral differentiation.

Main Results:

• Successful differentiation of multiple groups of EEL spectra representing specific fine structures across various experimental EELS-SI datasets.
• Enabling the investigation of particular fine structures by averaging spectra within identified clusters.
• Revealing the spatial distributions of different fine structures within materials.

Conclusions:

• The proposed machine learning strategy offers an efficient and data-driven method for analyzing EELS-SI data.
• This approach allows for detailed characterization of material microstructures based on spectral analysis.
• The method is broadly applicable to any hyperspectral imaging dataset without requiring prior information.