MDGAN is a generative adversarial network to generate materials microstructure images for materials discovery. The networks are trained on synthetic 2D microstructure images.
This software requires the input in the form of a numpy array of shapre (x, 128, 128) where x is the number of microstructure images and the dimension of microstructure is 128 by 128. The software takes the microstructure images as input to train the generative adversarial networks. The trained generator can then be used to produce microstructure images. The detailed description about data preprocessing and model can be found in the published paper given below.
- Python 2.7.12
- Numpy 1.14.0
- Keras 2.1.2
- Pickle 2.0
- TensorFlow 1.5.0-rc1
- H5PY 2.7.1
gan_training.py: The script that trains the GAN that can generate two-phase microstructure image. After training, the trained model will be saved in 'model' folder generated in the current repository.example_data.pkl: Example data of synthetic microstructure images, including 10 2D microstructures.
- Download the file containing the weights of VGG16, vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5, before the training.
- Run commend below, which trains the GAN and save the trained model in 'model' folder generated in the current repository.
python gan_training.py
The Rigorous Couple Wave Analysis simulation is supported by Prof. Cheng Sun's lab at Northwestern University. This work is primarily supported by the Center of Hierarchical Materials Design (NIST CHiMaD 70NANB14H012) and Predictive Science and Engineering Design Cluster (PS&ED, Northwestern University). Partial support from NSF awards DMREF-1818574, DMREF-1729743, DIBBS-1640840, CCF-1409601; DOE awards DE-SC0007456, DE-SC0014330; AFOSR award FA9550-12-1-0458; and Northwestern Data Science Initiative is also acknowledged.
Z. Yang, X. Li, L. C. Brinson, A. Choudhary, W. Chen, and A. Agrawal, “Microstructural Materials Design via Deep Adversarial Learning Methodology,” Journal of Mechanical Design, vol. 140, no. 11, p. 10, 2018.
- Zijiang Yang [email protected]
- Ankit Agrawal [email protected]