Machine learning enhanced analysis of EBSD data for texture representation

Abstract Generating reduced-order, synthetic grain structure datasets that accurately represent the measured grain structure of a material is important for reducing the cost and increasing the accuracy of computational crystal plasticity efforts. This study introduces a machine-learning-based approach, termed texture adaptive clustering and sampling (TACS), for generating representative Euler angle datasets that accurately mimic the crystallographic texture. The TACS approach employs K-means clustering and density-based sampling in a closed-loop iteration to create representative Euler angle datasets. Proof-of-principle experiments were performed on rolled and recrystallized low-carbon steel. Validation of the TACS approach was extended to twenty-two datasets, varying lattice structures, and complex crystallographic textures, thereby encompassing a broad range of materials and crystal structures. Kolmogorov-Smirnov (K-S) test comparisons underscore the performance of the TACS approach over traditional electron backscatter diffraction EBSD dataset reduction techniques, with average K-S test scores nearing 0.9, indicating a high-fidelity representation of the original datasets. In contrast, conventional methods display scores below 0.3, indicating less reliability of the structure representation. The independence of the TACS approach from material texture and its capability to autonomously generate datasets with predetermined data points demonstrates its unbiased potential in streamlining dataset preparation for crystallographic analysis.