Related Experiment Video
Updated: Jul 9, 2025
09:13
Characterization of Ultra-fine Grained and Nanocrystalline Materials Using Transmission Kikuchi Diffraction
Published on: April 1, 2017
13.7K
A lightweight transformer for faster and robust EBSD data collection
Harry Dong1, Sean Donegan2, Megna Shah2
1Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, 15289, USA. harryd@andrew.cmu.edu.
Scientific Reports
|December 1, 2023
Summary
This study introduces a novel transformer-based method to recover missing data slices in 3D electron back-scattered diffraction (EBSD) microscopy. The approach enhances data quality and collection speed for materials science applications.
Area of Science:
- Materials Science
- Data Science
- Microscopy
Background:
- Three dimensional electron back-scattered diffraction (3D EBSD) microscopy is vital in materials science.
- Data quality in 3D EBSD can be compromised during serial-sectioning collection.
- The sequential nature of 3D EBSD data lends itself to deep learning approaches.
Purpose of the Study:
- To develop a robust method for recovering missing slices in 3D EBSD volumes.
- To accelerate 3D EBSD data collection and improve overall data quality.
- To overcome challenges of deep learning with high-dimensional, scarce data.
Main Methods:
- A two-step method combining an efficient transformer model and a projection algorithm.
- Self-supervised learning using synthetically generated 3D EBSD data.
- Training a deep learning model to process sequential 3D EBSD data.
Main Results:
- Superior recovery accuracy on real 3D EBSD data compared to existing methods.
- Demonstrated effectiveness of the transformer model in processing sequential data.
- Successful training of a deep learning model using only synthetic data.
Conclusions:
- The proposed transformer-based method effectively recovers missing slices in 3D EBSD data.
- This approach enhances the robustness and efficiency of 3D EBSD data collection.
- Self-supervised learning with synthetic data is a viable strategy for training deep learning models in this domain.
