Dataset used in "Neural Networks for Topology Optimization" [link][pdf][code]

Download (~3 Gb) Yandex Drive

The dataset of topology optimization process. It contains the precise solutions of 10,000 randomly stated problems. Each object is a tensor of shape (100, 40, 40): 100 iterations, 40×40 grid.

We used Topy for the generation of the current dataset. It is based on SIMP approach.

The constraints and loads for each of the problem are chosen in the following way:

• The number of nodes with fixed x and y translations and the number of loads are sampled from the Poisson distribution:

• The nodes for each of the above described constraints are sampled from the distribution defined on the grid. The probability to choose the boundary node is 100 times higher than that for inner node.

• The load values are chosen as -1
• The volume fraction is sampled from the Gaussian distribution with mean of 0.5 and std of 0.1

You can generate your own dataset just by using scripts from code/ folder. sampler.py defines all the required distributions, and generate_data.py is just a convenient wrapper for ToPy. Install ToPy and run:

python code/generate_data.py --dir DIR_TO_SAVE --num NUMBER_OF_SAMPLES

Each tensor is stored in compressed format .npz. It could be extracted easily with numpy:

import numpy as np

data = np.load(PATH_TO_FILE)['arr_0']

We recommend to use horizontal and vertical flips as well as 90° rotation to augment the initial dataset. These operations allow one to get the dataset of 80,000 objects.

Please cite our work if you write a scientific paper using this code and/or dataset.

@article{sosnovik2017neural,
  title={Neural networks for topology optimization},
  author={Sosnovik, Ivan and Oseledets, Ivan},
  journal={arXiv preprint arXiv:1709.09578},
  year={2017}
}