Multi-modal Dataset of a Polycrystalline Metallic Material: 3D Microstructure and Deformation Fields

Abstract

The development of high-fidelity mechanical property prediction models for the design of polycrystalline materials relies on large volumes of microstructural feature data. Concurrently, at these same scales, the deformation fields that develop during mechanical loading can be highly heterogeneous. Spatially correlated measurements of 3D microstructure and the ensuing deformation fields at the micro-scale would provide highly valuable insight into the relationship between microstructure and macroscopic mechanical response. They would also provide direct validation for numerical simulations that can guide and speed up the design of new materials and microstructures. However, to date, such data have been rare. Here, a one-of-a-kind, multi-modal dataset is presented that combines recent state-of-the-art experimental developments in 3D tomography and high-resolution deformation field measurements.

A lightweight transformer for faster and robust EBSD data collection

“…2 . Due to limited real EBSD data, our goal is to train this model on a large and diverse synthetic dataset and demonstrate that it can generalize to real EBSD data, which we evaluate on two nickel superalloy EBSD volumes, one for alloy IN625 36 , 37 and one for alloy IN718 38 , 39 . The following repository contains the code for our method: https://github.com/hdong920/ebsd_slice_recovery .…”

“…Each dataset (both synthetic and experimental) includes orientation information at every voxel in a 3D image. For the experimental data, a substatial amount of preprocessing was done to handle the alignment of the data and clean up noise; a complete description of the preprocessing may be found in Chapman et al 2 and Stinville et al 38 In particular, we remove any grains smaller than 27 voxels ( ) and average orienations per grain. While operating on grain-average orientations simplifies the orientation prediction problem, it does not represent the real complexity of orientation fields, which often exhibit subtle local variations.…”

“…While not needed during training, these ID numbers will be used to smooth our model outputs and evaluate model accuracy. For experimental data, the IDs are found by segmenting the grains using a misorientation tolerance 2 , 38 , and for synthetic data, the IDs are generated alongside the orientation data. Boundaries: A volume of shape containing Boolean values indicating if a voxel is on the grain boundary.…”

“…3 contains example slices of (scaled and shifted) Cubochorics and Boundaries of IN625 and IN718.

Figure 3 Three consecutive crops of real EBSD slices from IN625 2 (left) and IN718 38 (right). The top row shows the cubochoric values that are scaled and shifted for better visualization, and the bottom row shows the same region in Boundaries.

…”

“…Now, we seek to understand how well our model transfers to real data by running our trained model on IN625 36 , 37 and IN718 38 , 39 . For both volumes, we subdivide each volume into nonoverlapping subvolumes such that all slices are of height and width 64 voxels.…”