This repo includes the dataset/scripts of 3D Convolutional Neural Network for material property prediction in paper:

Rao, Chengping, and Yang Liu. "Three-dimensional convolutional neural network (3D-CNN) for heterogeneous material homogenization." Computational Materials Science 184 (2020): 109850.

@article{rao2020three,
  title={Three-dimensional convolutional neural network (3D-CNN) for heterogeneous material homogenization},
  author={Rao, Chengping and Liu, Yang},
  journal={Computational Materials Science},
  volume={184},
  pages={109850},
  year={2020},
  publisher={Elsevier}
}

• training&plot.py: Data processing, training and postprocessing scripts ;
• data: The dataset for this application contains 2000 samples or (X, Y) pairs. X is the tensor of size 101x101x101 that carries the phase indication for a two-phase heterogeneous material microstructure (or Representative Volume Element). Y is a vector of size 12 that represents the effective material properties for the microstructure.
• saved_model: The model with best performance on validation set.
• gallery: Some figures.

Input X for the 3D Convolutional Neural Net, i.e. Cartisian grid carrying phase indication

Architecture of the 3D Convolutional Neural Net

Achieved prediction versus ground truth on testing data

• The code was developed based on TensorFlow 1.10.0 and Keras 2.2.4. All the runtime performance are evaluated on GeForce GTX 1080 Ti.