Comparison and validation of stochastic microstructure characterization and reconstruction: Machine learning vs. deep learning methodologies

In the world of computational materials science, the knowledge of microstructure is vital in understanding the process-microstructure–property linkage across various length-scales. To circumvent costly experimental characterizations, typically, analyses on ensembles of 3D microstructures within a numerical framework are preferred. Utilizing a moment invariants-based physical descriptor, the current work quantifies the variations in the microstructural topology of 3D synthetic data of polycrystalline materials. For the first time, the validation of synthetic microstructures based on two unique AI-based reconstruction approaches was compared, providing valuable insights into the diverse characteristics of each methodology. Virtual 3D microstructure volumes of forged Ti-7Al and additively manufactured 316L stainless steel alloys were generated from 2D experimental data using two methods — Markov Random Field (MRF) and deep learning-based volumetric texture synthesis. Quantitative evaluation and validation of the reconstructed volumes were carried out with the aid of moment invariants by comparing local features associated with grain-level properties, such as grain size and shape. The normalized central moments previously employed to compare 2D grain topology were expanded to 3D. With the advent of various reconstruction algorithms, especially AI-based, the validation methodology outlined in this work can be adopted to evaluate the robustness of various 3D reconstruction frameworks as well as ensure spatial equivalency of the target microstructures.