Graph pairs for GrainGNN

This dataset consists of a pickle file storing the graph pairs used to train GrainGraphGNN (often shortened to GrainGNN), a neural network for generating graph representations of directionally solidified grain structures. These grain structures are from simulations with thermal gradients aligned with the simulation’s Z direction, and the graph pairs consist of graph representations of the microstructure from data at Z = l and corresponding graph for the microstructure at Z = l +1. The graph consists of junction vertices, grain vertices, and edges. The features of junction vertices are the 3D position, thermal gradient magnitude, isotherm velocity in the Z direction, and the displacement of the vertices relative to their locations at Z = l – 1. The features of grain vertices are the 3D position of the grain’s associated vertices, grain cross-sectional area, grain excess volume above the interface height at Z = l, change in grain cross-sectional area between Z = l – 1 and Z = l, grain crystallographic orientation, and distance between the graph at Z = l and the corresponding graph at Z = l + 1. The features of edges are their lengths, assuming straight segments connecting grain vertices. More information about GrainGNN can be found in Qin et al.'s 2024 publication in the Journal of Computational Physics, available at https://doi.org/10.1016/j.jcp.2024.113061, and the GrainGNN source code and instructions on using this dataset for model training are available at https://github.com/YigongQin/GrainGraphNN. Using the code in the GrainGNN repository and the graph pairs in this repository, the neural network can be trained by running the following: python3 train.py --model_type=regressor --model_id=0 --device=cuda python3 train.py --model_type=classifier --model_id=1 --device=cuda