Bayesian Texture Optimization using Deep Neural Network-based Numerical Material Test

---

Description

• Bayesian Texture Optimization using Deep Neural Network-based Numerical Material Test (BayesTexOpt) project provides neural network (NN) structure (named DNN-3D), training parameters (weights and biases for the NN) and training datasets reported in the paper (R. Kamijyo et al., (2021) submitted).
• Using the DNN-3D, one can perform optimization of crystallographic texture in an aluminum alloy sheet to reduce in-plane anisotropy of Lankford value.
• The DNN-3D is constructed on TensorFlow and Keras.
• This repository also provides 3D-viewer to flexibly visualize the optization result.
• This project is related to Deep Neural Network-based Numerical Material Test (DNN-NMT) project.
• The detailed methodology can be found in the following publications.

Publications

1. R. Kamijyo, A. Ishii, and A. Yamanaka, Bayesian Texture Optimization using deep neural network-based numerical material test, International Journal of Mechanical Sciences, Vol. 223 (2022), 107285. https://doi.org/10.1016/j.ijmecsci.2022.107285

2. A. Yamanaka, R. Kamijyo, K. Koenuma, I. Watanabe and T. Kuwabara, "Deep neural network approach to estimate biaxial stress-strain curves of sheet metals", Materials & Design, Vol. 195 (2020), 108970. https://doi.org/10.1016/j.matdes.2020.108970

3. K. Koenuma, A. Yamanaka, I. Watanabe and T. Kuwabara, "Estimation of texture-dependent stress－strain curve and r-value of aluminum alloy sheet using deep learning", Materials Transactions, Vol. 61 (2020), pp. 2276-2283 https://doi.org/10.2320/matertrans.P-M2020853.

Requirements

• Anaconda3
• Some python libraries required for the BayesTexOpt can be installed by executing the batch file:

tf_env

Usage

Training DNN

1. Download the training data from Website of Yamanaka Laboratory, TUAT.
2. Save the training data to any directory. For example, 'E:/'.
3. Edit "common/rawdata.py" so that ROOT_DIR is the same directory as that where the training data is saved.

ROOT_DIR = 'E:/'

4. Run "dataset.py" to prepare the training and validation datasets.

conda activate tf_env
python dataset.py

5. Run "train_tfmodel.py" to train the DNN-3D using the training dataset. The trained DNN-3D will be saved in the directroy "tf_models/dnn3d/model.py".

python train_tfmodel.py

6. Run "draw_rvalue.py" to estimate Lankford values using the trained DNN-3D.

Texture optimization using the trained DNN-3D

1. Perform "Bayesian Texture Optimization" by running "optimize_rvalue_BO.py". The optimization results are saved to the file named "Opt_result/ev_all.dat".
2. The parameters of Kernel function used in the Gaussian process regression (GPR) are saved in the file named "Opt_result/model.dat".
3. The optimization results (i.e., 3D distribution of predictive mean, standard deviation, and acquisition function) can be visualize by running "BayesTexOpt_visualizer.py" with an example data "example_data/ev_all_3d.dat". Afetr running the script, you can rotate three-dimensional graphs of mean function, standard deviation and acquisition function:

License

BSD License (3-clause BSD License)

Author

Yamanaka Research Group @ TUAT