Robust Spectral Anomaly Detection in EELS Spectral Images via Three Dimensional Convolutional Variational Autoencoders

TL;DR

The paper tackles unsupervised detection of spectral anomalies in EELS-SI data by introducing a Three-Dimensional Convolutional Variational Autoencoder (3D-CVAE) that captures both spectral and spatial correlations via 3D convolutions. The learning objective combines a cross-entropy reconstruction term with a KL-divergence regularization, with a beta weight of 1.2 and a latent space of 40 dimensions. Anomalies are benchmarked by injecting Fe L-edge peak shifts of 2.5 eV within the 690-730 eV window, and performance is compared against PCA with 3-5 components; results show clear bimodal separation in reconstruction-based anomaly scores and robust detection across shift magnitudes. Latent-space analyses reveal that paired normal and anomalous sub-images map to nearby regions, and the method maintains high reconstruction fidelity in noisy regions, providing a practical unsupervised framework for spectral anomaly detection in EELS-SI data with potential diffusion-based denoising extensions.

Abstract

We introduce a Three-Dimensional Convolutional Variational Autoencoder (3D-CVAE) for automated anomaly detection in Electron Energy Loss Spectroscopy Spectrum Imaging (EELS-SI) data. Our approach leverages the full three-dimensional structure of EELS-SI data to detect subtle spectral anomalies while preserving both spatial and spectral correlations across the datacube. By employing negative log-likelihood loss and training on bulk spectra, the model learns to reconstruct bulk features characteristic of the defect-free material. In exploring methods for anomaly detection, we evaluated both our 3D-CVAE approach and Principal Component Analysis (PCA), testing their performance using Fe L-edge peak shifts designed to simulate material defects. Our results show that 3D-CVAE achieves superior anomaly detection and maintains consistent performance across various shift magnitudes. The method demonstrates clear bimodal separation between normal and anomalous spectra, enabling reliable classification. Further analysis verifies that lower dimensional representations are robust to anomalies in the data. While performance advantages over PCA diminish with decreasing anomaly concentration, our method maintains high reconstruction quality even in challenging, noise-dominated spectral regions. This approach provides a robust framework for unsupervised automated detection of spectral anomalies in EELS-SI data, particularly valuable for analyzing complex material systems.

Figures (7)

• Figure 1: Comparison of VAE and PCA reconstructions and their anomaly detection performance. (a) Split visualization of the EELS-SI datacube reconstruction, with VAE (left) and PCA (4 components, right) results shown as z-direction intensity sums. (b) VAE reconstruction error heatmap showing Pearson Correlation Coefficients between original and reconstructed spectra. (c) Corresponding PCA reconstruction error heatmap. In both (b) and (c), green circles indicate successfully detected anomalies using Otsu's thresholding method, while red circles mark undetected anomalous regions. Lower PCC values (lighter colors) indicate greater deviation between original and reconstructed spectra.
• Figure 2: Example of an injected peak shift anomaly in EELS spectra. The original Fe L-edge spectrum (black) compared to an artificially introduced 2.5 eV peak shift (red segment)
• Figure 3: Distribution of pixels across Pearson Correlation Coefficient (PCC) values for VAE (top) and PCA with 4 components (bottom). Each histogram shows the distribution of normal (bulk) and anomalous pixels on a logarithmic scale. PCC values range from 0.2 to 1.0, where 1.0 indicates perfect correlation between original and reconstructed spectra.The VAE shows clear bimodal separation between normal and anomalous distributions, enabling reliable anomaly detection, while PCA distributions remain overlapped.
• Figure 4: Performance comparison across different magnitudes of peak shifts. F1-scores for VAE (red) and PCA with 3 (blue), 4 (orange), and 5 (green) components. VAE maintains consistently high F1-scores across all shift magnitudes, while PCA exhibits periodic fluctuations in performance. PCA with 3 components shows the best performance among PCA variants, though its effectiveness varies with shift magnitude.
• Figure 5: Visualization of latent space relationships through cosine similarity between 40-dimensional encodings of EELS-SI sub-image pairs. Each point compares an unmodified image (Y axis) with its anomaly-injected counterpart (X axis). The diagonal values close to 1 demonstrate that the encoder places paired images in nearly identical positions in the latent space, confirming that our model successfully recognizes anomalous spectra as variants of their normal counterparts.