Topodiff is a conditional diffusion-model-based architecture to perform performance-aware and manufacturability-aware topology optimization that overcomes the issues of Generative Adversarial Networks (GANs) such as difficult to train, limited generalizability and neglecting manufacturability. It introduces a surrogate model-based guidance strategy that actively favors structures with low compliance and good manufacturability.
The dataset for the diffusion, regression, and classifier models can be downloaded from here
Download the dataset and set the path to the dataset in the config file.
2D topology structures could be generated by Topodiff conditioned on the boundary and loading conditions. A few examples are shown below:
Install the requirements using:
pip install -r requirements.txtBefore training the model, take a loo at the Dataset Configuration & Paths section of the config file
and set the correct paths to the dataset.
Next, run the following commands to train the diffusion model, classifier model for floating material, and the and regressor model for compliance:
python train.py
python train_classifier.py
python train_regressor.pyBy default, the generated topologies are conditioned on the boundary and loading conditions that have not been seen in the training process. Run the following command to generate topologies:
python inference.pyTo cite this work, please use the following reference:
@inproceedings{maze2023diffusion,
title={Diffusion models beat gans on topology optimization},
author={Maz{\'e}, Fran{\c{c}}ois and Ahmed, Faez},
booktitle={Proceedings of the AAAI conference on artificial intelligence},
volume={37},
number={8},
pages={9108--9116},
year={2023}
}