[論文レビュー] Multi-modal data-driven microstructure characterization

Qi Zhang, Santiago Benito+2 more|arXiv (Cornell University)|Jan 18, 2026

Machine Learning in Materials Science被引用数 0

ひとこと要約

The paper presents an autonomous, data-driven workflow for EBSD microstructure analysis by combining PCA, constrained non-negative matrix factorization, and a variational autoencoder to achieve grain/phase segmentation and boundary detection without user input.

ABSTRACT

Electron backscatter diffraction is one of the most prevalent techniques used for microstructural characterization. In recent years, there has been an increase in the use of data-driven methods to analyze raw Kikuchi patterns. However, most of these require user input and the interpretation of the data-derived features is often challenging and subject to \textit{informed interpretation}. By using a combination of principal component analysis, constrained non-negative matrix factorization, and a variational autoencoder along with information-theoretical considerations on a multimodal dataset, it is shown that a) automated decision on method-specific hyperparameters, here the number of components in principal component analysis, the number of components for constrained non-negative matrix factorization, and the selection of reference constraints; and b) latent space features can be mapped to physically-meaningful quantities. In addition, the recommended region-of-interest (ROI) size for optimal model performance is approximated automatically to be twice the characteristic grain size based on information content of the dataset. Implemented in a workflow, this allows for a transferable, dataset-specific autonomous data-driven phase and grain segmentation including grain boundary detection and the analysis of very-small-angle intra-grain variations to complement conventional electron backscatter analysis.

研究の動機と目的

• Automate hyperparameter decisions for PCA and cNMF (e.g., number of components, reference constraints) using information-theoretic criteria.
• Map latent space features to physically meaningful microstructure quantities.
• Develop a transferable workflow for automated phase and grain segmentation including grain boundary detection.
• Enable analysis of intra-grain orientation variations and defects beyond conventional EBSD.
• Apply the workflow to a partially-reduced iron ore pellet to demonstrate transferability and robustness.

提案手法

• Preprocess EBSD data and discretize the region of interest (ROI) for subsequent analysis.
• Use PCA (or IPCA) to reduce dimensionality and automatically determine the optimal number of components via a knee-point detection on the cumulative variance curve.
• Cluster PCA representations with a Gaussian Mixture Model and automatically select the number of clusters using a composite AIC/BIC/SC score.
• Perform constrained non-negative matrix factorization (cNMF) using data-driven reference components automatically selected from Mahalanobis-distance-based analysis.
• Apply a convolutional variational autoencoder (VAE) to learn latent representations of Kikuchi patterns and relate latent features to physical/EDS data.
• Define anomalies via Mahalanobis distance and chi-square thresholds to identify grain/phase boundaries and heterogeneities.

実験結果

リサーチクエスチョン

• RQ1 Can hyperparameters for PCA and cNMF be automatically determined from the data without user input?
• RQ2 How can latent-space features from PCA, cNMF, and VAE be correlated with physically meaningful microstructural quantities (phases, boundaries, defects)?
• RQ3 What ROI size yields optimal performance for data-driven microstructure segmentation and boundary detection?
• RQ4 Are grain and phase boundaries identifiable from data-driven weight maps and anomaly analysis, and how do they compare to conventional EBSD results?
• RQ5 To what extent can a fully autonomous workflow delineate intra-grain orientation variations and substructures?

主な発見

• An automated, transferable data-driven workflow for EBSD microstructure characterization is feasible by integrating PCA, GMM clustering, Mahalanobis-distance anomalies, cNMF, and VAE.
• The method enables autonomous decision on hyperparameters and automatic selection of reference constraints for cNMF based on data distribution.
• Anomalies (via Mahalanobis distance and chi-square threshold) often coincide with grain or phase boundaries and intra-grain heterogeneities.
• ROI size is linked to an optimal length scale, approximately twice the characteristic grain size, improving segmentation robustness.
• VAE latent features correlate with EDS elemental distributions and band contrast, providing a complementary unsupervised representation of microstructure.
• The approach yields consistent grain segmentation and boundary detection and complements conventional EBSD analysis on a partially-reduced iron ore pellet.