SLEDO (Sequential Learning Engineering Design Optimiser) is a tool for optimising parametric component designs, developed for fusion engineering.
sledo uses Bayesian Optimisation to intelligently select candidate designs from a parametric search space and MOOSE to evaluate the design performance via scalable multiphysics simulation.
We recommend using a python virtual environment to install sledo.
Run the following to create and activate a virtual environment in your current working directory.
Skip this step if you intend to install sledo into an existing python environment.
python3 -m venv venv
source venv/bin/activate
From the sledo root directory (the directory containing this README), ensure your virtual environment is activated and then run:
pip install .
If you intend to use the jupyter notebook tutorials and are using a virtual environment, also run the following line (replacing `venv`` with the name of your virtual environment if it is different):
# Enable virtual environment in the tutorial notebooks.
python3 -m ipykernel install --user --name=venv
sledo can now be imported into python as:
import sledo
To test the installation, run python3 -c "import sledo". If the installation was unsuccessful, an error will be raised.
Developers may wish to create an editable installation. This allows changes to the source code to immediately take effect without the need to re-install sledo. This can be useful when running tests and other scripts.
To install sledo this way, use the --editable flag when running the pip install line as follows:
# Install as an editable installation.
pip3 install --editable .
To start using sledo, please refer to the tutorials and examples folders for example usage.
Before using sledo, please see the following section on providing access to your MOOSE application of choice for running simulations.
While sledo itself is a pure python project, it is intended to be used alongside MOOSE.
The simplest way to give sledo visibility of your MOOSE installation and chosen MOOSE app is to fill in the required paths in moose_config.json.
If you don't have MOOSE installed already, there are several MOOSE applications available alongside sledo in the Aurora Multiphysics organisation (https://github.com/aurora-multiphysics), such as proteus, aurora, and apollo; each is focussed on different physics domains, so choose the one appropriate for your design problem.
Luke Humphrey, UK Atomic Energy Authority
Aleksander Dubas, UK Atomic Energy Authority
Andrew Davis, UK Atomic Energy Authority
Lloyd Fletcher, UK Atomic Energy Authority
If using sledo in your research, please cite the following:
-
L R Humphrey et al 2024 Plasma Phys. Control.Fusion 66 025002
Bibtex:
@article{humphrey_machine_2024, title = {Machine learning techniques for sequential learning engineering design optimisation}, volume = {66}, issn = {0741-3335, 1361-6587}, url = {https://iopscience.iop.org/article/10.1088/1361-6587/ad11fb}, doi = {10.1088/1361-6587/ad11fb}, language = {en}, number = {2}, urldate = {2023-12-21}, journal = {Plasma Physics and Controlled Fusion}, author = {Humphrey, L R and Dubas, A J and Fletcher, L C and Davis, A}, month = feb, year = {2024}, pages = {025002}, }
Please also consider citing sledo's dependencies as indicated on their individual websites:
- MOOSE (https://mooseframework.inl.gov/citing.html)
- Ray-Tune (https://docs.ray.io/en/latest/tune/index.html#citing-tune)
- BoTorch (https://botorch.org/docs/papers)
- NumPy (https://numpy.org/citing-numpy/)
- SciPy (https://scipy.org/citing-scipy/)
- Matplotlib (https://matplotlib.org/stable/users/project/citing.html)