NTopo: Mesh-free Topology Optimization using Implicit Neural Representations

Paper on arxiv

This repository is the official implementation of NTopo. For the purpose of understandability, this is a reimplementation of the ideas which were used to generate the results in the paper. NTopo is mesh-free topology optimization solver using neural representations.

Comparisons with FEM were generated using a variation of the code topopt_cholmod available from DTU.

Running the code

Requirements

This code has been tested with tensorflow 2.3, python 3.7 and CUDA 10.1 .

The complete list of packages required is shown in requirements.txt. After setting up a python environment and installing the appropriate tensorflow requirements such as CUDA, the packages can be installed using

pip install -r requirements.txt

To see the usage of the main program run the command

python run.py --help

Training

Running

python run.py list_problems

will list the available problems that are not related to solution spaces

python run.py train Beam2D

generates a config file and runs it. Results will be stored in a subfolder ./results/<experiment>.

A modified config file can be run with

python run.py train_config <config_file>

Evaluation

To run the evaluation, one can run

python run.py evaluate <folder> <density_weights>

In practice, this could look something like

python run.py evaluate results/Beam2DSpaceVolume-Adam-###_vol_#.# density_model-000100 --n_q_samples=100

which will create density images and a data.json file in a new results folder.

Training a solution space

Running

python run.py list_problems_space

will list problems related to solution spaces. Similarly, other commands are available with a _space suffix to train and evaluate models.

License

See the LICENSE file for license rights and limitations.