Graph neural networks for efficient learning of mechanical properties of polycrystals

This repository holds the data and code required to reproduce results presented in "Graph neural networks for efficient learning of mechanical properties of polycrystals".

@article{HESTROFFER2023,
title = {Graph neural networks for efficient learning of mechanical properties of polycrystals},
journal = {Computational Materials Science},
volume = {217},
pages = {111894},
year = {2023},
issn = {0927-0256},
doi = {https://doi.org/10.1016/j.commatsci.2022.111894},
url = {https://www.sciencedirect.com/science/article/pii/S092702562200605X},
author = {Jonathan M. Hestroffer and Marie-Agathe Charpagne and Marat I. Latypov and Irene J. Beyerlein},
}

Setup

Install dependencies. Please note that installing PyTorch and PyTorchGeometric will be more involved and custom to your machine/GPU. See links provided.

pip install -r requirements.txt

Create Graphs

Generate microstructure graphs for all representative volume elements (RVEs).

python create_graphs.py

Prepare Data

Assemble graphs and rve mechanical response into PyTorch datalists.

python write_data.py

Run Model Evaluations

The different evaluations presented in the original paper are:

1. 10-fold cross-validation of texture groups (A-G)
2. Train (A-G) / Test (A-G)
3. Train (A-G) / Test (H-L)
4. 5-fold reduced data 'psuedo cross-validation', Train (A-G) / Test (H-L)

For each evaluation, loss histories, parity plots, and model checkpoints are outputted.

Usage: python model.py [OPTIONS]
Options:
  --eval INT                   Evaluation number (e.g., 1 - 4, default: 1)
  --prop STRING                Material property of interest (stiffness/strength, default: stiffness)
  --config INT                 Hyper-parameter configuration number (default: 0)
  --config_dir PATH            Directory with hyper-parameter configuration .jsons (default: ./config/)
  --output_dir PATH            Output directory (default: ./config/)
  --seed INT                   Random seed (default: 42)