In this repository, we run an investigation of Multi-fidelity Bayesian Optimization (MFBO) methods and their application to chemistry and materials problems.
We first conduct some benchmarks with 2 different acquisition functions and 2 synthetic functions. We also study the parameters favoring the MF approach over the standard single fidelity (SFBO).
We run 3 benchmarks in the chemical domain (optimization of Covalent Organic Frameworks (COFs), polarizable molecules and molecule solvation energy). Our study shows how MFBO can effectively reduce the overall cost of optimization tasks and provide guidelines for using this method.
📄 The corresponding paper can be found here
Create a new environment
python -m venv venv
Activate the environment and install the package
source venv/bin/activate
pip install .
If you want to use the hydra SLURM plugin to launch jobs on a cluster
pip install hydra-submitit-launcher --upgrade
If you want to launch the synthetic functions benchmark, use this command
python src/chem_mfbo/benchmark/benchmark.py
Additionally, the sweep over synthetic functions can be run via
python src/chem_mfbo/benchmark/benchmark.py --config-name=synthetic_sweep.yaml
The benchmarks for chemistry and materials design can be reproduced using the corresponding .sh file
./chemistry_benchmarks.sh
To plot the benchmarks results, include the corresponding results path in the path option of the corresponding plotting config file included in config_plots/ and run the associated plotting script. As an example, the synthetic functions benchmark can be run with:
python src/chem_mfbo/benchmark/plot_synthetic.py
@misc{sabanzagil2024bestpracticesmultifidelitybayesian,
title={Best Practices for Multi-Fidelity Bayesian Optimization in Materials and Molecular Research},
author={Víctor Sabanza-Gil and Riccardo Barbano and Daniel Pacheco Gutiérrez and Jeremy S. Luterbacher and José Miguel Hernández-Lobato and Philippe Schwaller and Loïc Roch},
year={2024},
eprint={2410.00544},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2410.00544},
}
This work was created in a collaboration between Atinary Technologies Inc. and LIAC - LPDC, as part of the EPFLglobaLeaders program. The project was funded by NCCR Catalysis (grant number 180544), a National Centre of Competence in Research funded by the Swiss National Science Foundation, and by the the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement N° 945363.