Robust Spectral Anomaly Detection in EELS Spectral Images via 3D Convolutional Variational Autoencoders

Area of Science:

• Materials Science
• Spectroscopy
• Machine Learning

Background:

• Electron energy-loss spectroscopy spectrum imaging (EELS-SI) generates complex 3D datacubes.
• Detecting subtle spectral anomalies in EELS-SI data is crucial for material analysis.
• Existing methods may struggle with preserving spatial and spectral correlations.

Purpose of the Study:

• To introduce and evaluate a 3D Convolutional Variational Autoencoder (3D-CVAE) for automated anomaly detection in EELS-SI data.
• To compare the performance of 3D-CVAE against Principal Component Analysis (PCA) for anomaly detection.
• To establish a robust framework for unsupervised spectral anomaly identification in complex materials.

Main Methods:

• Development of a 3D-CVAE model leveraging the full 3D structure of EELS-SI data.
• Training the 3D-CVAE on bulk spectra to reconstruct defect-free material features using cross-entropy loss.
• Simulating material defects using Fe L-edge ΔE peak shifts for performance evaluation.
• Comparative analysis with Principal Component Analysis (PCA).

Main Results:

• 3D-CVAE demonstrated superior anomaly detection performance compared to PCA, maintaining consistency across varying defect magnitudes.
• The method achieved clear bimodal separation between bulk and anomalous spectra, enabling reliable classification.
• Lower-dimensional representations derived from 3D-CVAE proved robust to anomalies.
• High reconstruction quality was maintained even in noise-dominated spectral regions.

Conclusions:

• The 3D-CVAE offers a robust, unsupervised framework for automated spectral anomaly detection in EELS-SI data.
• This approach is particularly valuable for analyzing complex material systems where subtle defects are prevalent.
• The model effectively preserves spatial and spectral correlations, enhancing detection accuracy.