[Paper Review] Advances in electron backscatter diffraction

Alex Foden, Alessandro Previero+1 more|arXiv (Cornell University)|Aug 12, 2019

X-ray Diffraction in Crystallography2 references4 citations

TL;DR

This paper advances electron backscatter diffraction (EBSD) through open-source algorithms and machine learning, enabling high-precision orientation mapping via unsupervised PCA and multivariate statistics, improved indexing with refined template matching, and phase classification using transfer learning with convolutional neural networks. It demonstrates a high-quality experimental reference diffraction sphere via direct electron detector and dynamical simulation, significantly enhancing microstructural analysis accuracy.

ABSTRACT

We present a few recent developments in the field of electron backscatter diffraction (EBSD). We highlight how open source algorithms and open data formats can be used to rapidly to develop microstructural insight of materials. We include use of AstroEBSD for single pixel based EBSD mapping and conventional orientation mapping; followed by an unsupervised machine learning approach using principal component analysis and multivariate statistics combined with a refined template matching method to rapidly index orientation data with high precision. Next, we compare a diffraction pattern captured using direct electron detector with a dynamical simulation and project this to create a high quality experimental "reference diffraction sphere". Finally, we classify phases using supervised machine learning with transfer learning and a convolutional neural network.

Motivation & Objective

• To accelerate microstructural insight in materials science through open-source EBSD algorithms and data formats.
• To improve orientation mapping precision and speed using unsupervised machine learning and refined template matching.
• To generate a high-fidelity experimental reference diffraction sphere by projecting dynamical simulations onto direct electron detector data.
• To enable accurate phase classification using supervised machine learning with transfer learning and convolutional neural networks.
• To integrate these innovations into a cohesive workflow for 2D, 3D, and 4D EBSD analysis.

Proposed method

• Utilizes AstroEBSD for single-pixel and conventional EBSD mapping to enable flexible data acquisition.
• Applies principal component analysis and multivariate statistics to reduce dimensionality and enhance indexing efficiency.
• Implements a refined template matching method to improve indexing precision in orientation mapping.
• Projects a dynamical simulation of a diffraction pattern onto experimental data from a direct electron detector to create a high-quality reference diffraction sphere.
• Employs transfer learning with a pre-trained convolutional neural network to classify phases from EBSD patterns with minimal labeled data.
• Combines open data formats and open-source software to ensure reproducibility and community adoption.

Experimental results

Research questions

• RQ1 How can open-source algorithms and data formats accelerate the development of microstructural analysis in EBSD?
• RQ2 To what extent can unsupervised machine learning improve indexing speed and precision in EBSD orientation mapping?
• RQ3 Can a dynamically simulated diffraction pattern projected onto experimental data produce a high-fidelity reference diffraction sphere?
• RQ4 How effective is transfer learning with convolutional neural networks for phase classification in EBSD data?
• RQ5 What is the impact of integrating these methods into a unified, open workflow for 4D EBSD analysis?

Key findings

• The integration of unsupervised machine learning with multivariate statistics and refined template matching significantly improves indexing precision and speed in EBSD data.
• A high-quality experimental reference diffraction sphere was successfully generated by projecting a dynamical simulation onto direct electron detector data, enhancing indexing accuracy.
• Phase classification using transfer learning with a convolutional neural network achieved high accuracy with limited labeled training data.
• The use of open-source tools like AstroEBSD and standardized data formats enables rapid, reproducible, and community-driven EBSD analysis.
• The combined approach enables high-precision 4D EBSD mapping, supporting advanced microstructural characterization in materials science.
• The workflow demonstrates strong potential for adoption in both academic and industrial materials research settings.