Application of deep learning for identification of grain boundaries in atomistic simulation data

Grain boundaries (GB) characteristics, such as tilt, twist, ∑3, ∑5, etc., play a critical role in determining the microstructural and functional properties of the materials. Due to experimental challenges, atomistic simulations have become a powerful alternative for studying these grain boundaries. However, the post-processing and interpretation of simulation data often require expert input, which can introduce subjectivity into the analysis. This study proposes a deep learning (DL)-based framework for the automated segmentation of grain boundary atoms from bulk atoms in three-dimensional atomic configurations derived from molecular dynamics (MD) simulations. We utilize a publicly available MD simulation dataset for aluminum GB structures. Raw MD simulation snapshots from this dataset are pre-processed to extract features representative of diverse grain boundary configurations. We employ K-means clustering to explore the structural classes within the dataset, while a 3D CNN model is later used for classifying bulk and grain boundary atoms to identify grain boundaries. The K-means clustering analysis revealed distinct patterns in the dataset for grain boundary and bulk atoms. The 3D CNN model demonstrates classification accuracy comparable to that of the commonly used common neighbor analysis (CNA) method in identifying grain boundary atoms across a range of misorientation angles. This work highlights the potential of DL-based approaches for consistent and accurate characterization of grain boundaries in atomistic simulations.