Strategies for EELS Data Analysis. Introducing UMAP and HDBSCAN for Dimensionality Reduction and Clustering

Area of Science:

• Materials Science
• Spectroscopy
• Data Analysis

Background:

• Electron Energy Loss Spectroscopy (EELS) generates complex spectrum images.
• Traditional segmentation methods for EELS data can be limited.
• Advanced algorithms offer potential for improved analysis.

Purpose of the Study:

• To evaluate the effectiveness of HDBSCAN and UMAP for EELS spectrum image segmentation.
• To compare these novel methods against existing EELS clustering approaches.
• To demonstrate the application of these algorithms on real experimental data.

Main Methods:

• Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) for clustering.
• Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction.
• Application to both synthetic and experimental core-loss EELS spectrum images.

Main Results:

• HDBSCAN and UMAP demonstrated superior performance in segmenting EELS spectrum images compared to conventional methods.
• Analysis of a synthetic dataset confirmed the enhanced results from UMAP and HDBSCAN.
• Successful application on a core-shell nanoparticle dataset, highlighting complementary insights from different algorithm combinations.

Conclusions:

• UMAP and HDBSCAN represent state-of-the-art approaches for EELS data analysis.
• The combined use of different algorithms, such as Nonnegative Matrix Factorization (NMF) with UMAP and HDBSCAN, provides a more comprehensive understanding of complex datasets.
• These advanced methods offer significant improvements for EELS spectrum image segmentation and analysis.