Data-driven approach for synchrotron X-ray Laue microdiffraction scan analysis

Abstract

A novel data‐driven approach is proposed for analyzing synchrotron Laue X‐ray microdiffraction scans based on machine learning algorithms. The basic architecture and major components of the method are formulated mathematically. It is demonstrated through typical examples including polycrystalline BaTiO3, multiphase transforming alloys and finely twinned martensite. The computational pipeline is implemented for beamline 12.3.2 at the Advanced Light Source, Lawrence Berkeley National Laboratory. The conventional analytical pathway for X‐ray diffraction scans is based on a slow pattern‐by‐pattern crystal indexing process. This work provides a new way for analyzing X‐ray diffraction 2D patterns, independent of the indexing process, and motivates further studies of X‐ray diffraction patterns from the machine learning perspective for the development of suitable feature extraction, clustering and labeling algorithms.

Processing Laue Microdiffraction Raster Scanning Patterns with Machine Learning Algorithms: A Case Study with a Fatigued Polycrystalline Sample

“…We applied the PCA to the data matrix M with Scikit-Learn [37], and displayed the variances of PCs in descending order in Figure 3, in which the variances of the first 64 PCs were shown in the inset. The distribution was highly skewed but the skewness was not as significant as in the work of Song et al [31], therefore more PCs needed to be truncated. To sum up, the pipeline for feature extraction consisted of (i) pixel agglomeration with HAC algorithm and subsequently (ii) factorization of data matrix via PCA.…”

“…In analogy to the preceding work [31,32,35,38], the treatment consisted of two steps: (i) feature extraction and (ii) grid segmentation. For pragmatic consideration, we adopted unsupervised machine learning algorithms since the training data might not be always available.…”

“…Then, since the background occupying the majority of Laue pattern contained nothing other than noise, a moderate resolution coarsening could speed up the image processing and smooth the noise [31]. In this regard, the normalized images were compressed to 512 × 512 pixels by 2 × 2 binning (averaging).…”

“…A number of algorithms serve to extract features for further grid segmentation. Song et al [31] applied CNN to extract latent features. However, at present, the authors did not have sufficient patterns to train the CNN, therefore unsupervised machine learning algorithms that did not require training dataset were employed herein to extract latent features of each pattern.…”

“…Zhou et al [30] proposed using the distribution of average recorded intensities and average filtered intensities of the raster scanned diffraction patterns to visualize the characteristics of microstructural features. Recent application of convolutional neural networks (CNN) to diffraction patterns has been involved with the extraction of features for further clustering and labeling the raster scanned diffraction patterns [31] or for the identification of crystal structures [32].…”