This repository contains code to train an 3D axial transformer to recover missing EBSD slices, as described in "A lightweight transformer for faster and robust EBSD data collection". The src folder contains code files for the model, projection algorithm, and training procedure. Due to the size of the trained model and data files, they can be found in this Google Drive. The naming convention for the data files is train_mu<mean grain size>_sig0-4_trans<mean transformations per grain>.dream3d for training volumes and val_mu<mean grain size>_sig0-4_trans<mean transformations per grain>_<volume number>.dream3d for validation volumes.
Described more in detail in the paper, the entire procedure is a two-step process:
-
The transformer takes an ESBD volume in cubochoric values (the shared model was trained only on volumes with one missing slice) and produces a cubochoric volume of the same shape as the input.
-
The projection algorithm takes the output corresponding to the missing slice and adjacent slices to predict grain IDs for each voxel in the missing slice.
If you use or build upon this work, please cite our paper:
@article{dong2023lightweight,
title={A lightweight transformer for faster and robust EBSD data collection},
author={Dong, Harry and Donegan, Sean and Shah, Megna and Chi, Yuejie},
journal={Scientific Reports},
volume={13},
number={1},
pages={21253},
year={2023},
publisher={Nature Publishing Group UK London}
}
This work was funded in part by the Data-Driven Discovery of Optimized Multifunctional Material Systems Center of Excellence (D3OM2S CoE), contract number FA8650-19-2-5209.