PaDGAN

Experiment code associated with the paper: PaDGAN: A Generative Adversarial Network for Performance Augmented Diverse Designs

Please also check our more recent work MO-PaDGAN for Design Reparameterization and Optimization, where we addressed the multi-objective scenario and demonstrated the method's effectiveness in multi-objective design optimization.

Check out my IDETC 2020 presentation here.

License

This code is licensed under the MIT license. Feel free to use all or portions for your research or related projects so long as you provide the following citation information:

Chen, W., and Ahmed, F. (November 10, 2020). "PaDGAN: Learning to Generate High-Quality Novel Designs." ASME. J. Mech. Des. March 2021; 143(3): 031703.

@article{chen2020padgan,
  title={PaDGAN: Learning to Generate High-Quality Novel Designs},
  author={Chen, Wei and Ahmed, Faez},
  journal={Journal of Mechanical Design},
  volume={143},
  number={3},
  publisher={American Society of Mechanical Engineers Digital Collection}
}

Required packages

• tensorflow < 2.0.0
• sklearn
• numpy
• matplotlib
• seaborn

Usage

Synthetic examples

1. Go to example directory:

   cd synthetic

2. Run single experiment:

   python run_experiment.py

   positional arguments:

   mode	train or evaluate
   data	dataset name (specified in datasets.py; available datasets are Ring2D, Grid2D, Donut2D, ThinDonut2D, and Arc2D)
   func	function name (specified in functions.py; available functions are Linear, MixGrid, MixRing, MixRing4, and MixRing6)
   

   optional arguments:

   -h, --help            	show this help message and exit
   --inf			maximize quality but not diversity
   --naive			use naive loss for quality
   --lambda0		coefficient controlling the weight of quality in the DPP kernel
   --lambda1		coefficient controlling the weight of the performance augmented DPP loss in the PaDGAN loss
   --disc_lr		learning rate for the discriminator
   --gen_lr		learning rate for the generator
   --id			experiment ID
   --batch_size		batch size
   --train_steps		training steps
   --save_interval 	number of intervals for saving the trained model and plotting results
   

   The default values of the optional arguments will be read from the file synthetic/config.ini.

   The trained model and the result plots will be saved under the directory synthetic/trained_gan/<data>_<func>_GAN_<lambda0>_<lambda1>/<id>, where <data>, <func>, <lambda0>, <lambda1>, and <id> are specified in the arguments or in synthetic/config.ini.

   Note that we can set lambda0 and lambda1 to zeros to train a vanilla GAN.

   Datasets and functions are defined in synthetic/functions.py and synthetic/datasets.py, respectively.

   Specifically, here are the dataset and function names (<data> and <func>) for the three synthetic examples in the paper:

   Example	Dataset name	Function name
   I	Grid2D	MixRing4
   II	Donut2D	MixRing6
   III	ThinDonut2D	MixRing6

Airfoil example

1. Install XFOIL.

2. Go to example directory:

   cd airfoil

3. Download the airfoil dataset here and extract the NPY files into airfoil/data/.

4. Go to the surrogate model directory:

   cd surrogate

5. Train a surrogate model to predict airfoil performances:

   python train_surrogate.py train

   positional arguments:

   mode	train or evaluate
   

   optional arguments:

   -h, --help            	show this help message and exit
   --save_interval		interval for saving checkpoints
   

6. Go back to example directory:

   cd ..

7. Run the experiment:

   python run_experiment.py train

   positional arguments:

   mode	train or evaluate
   

   optional arguments:

   -h, --help            	show this help message and exit
   --naive			use naive loss for quality
   --lambda0		coefficient controlling the weight of quality in the DPP kernel
   --lambda1		coefficient controlling the weight of the performance augmented DPP loss in the PaDGAN loss
   --id			experiment ID
   

   The default values of the optional arguments will be read from the file airfoil/config.ini.

   The trained model and the result plots will be saved under the directory airfoil/trained_gan/<lambda0>_<lambda1>/<id>, where <lambda0>, <lambda1>, and <id> are specified in the arguments or in airfoil/config.ini. Note that we can set lambda0 and lambda1 to zeros to train a vanilla GAN.

Please check out an example of using PaDGAN to address multivariate performance enhancement here.