This repository contains the implementation of the paper

Neural Thermodynamic Integration: Free Energies from Energy-based Diffusion Models by Bálint Máté, François Fleuret and Tristan Bereau.

The install.sh script will create a virtualenv necessary to run the experiments. The only requirement for this is python>=3.9.

The notebook toy_example.ipynb contains a simple experiment that demonstrates the idea on a 1D Gaussian mixture.

The run_exp.sh activates the virtualenv created by install.sh and then executes experiments/main.py using the configs experiments/config.yaml and experiments/LJ3D.yaml. When executing for the first time, it begins with generating the training data using MCMC. The samples are then dumped to files and loaded in later runs.

All the plots and metrics are also logged to the experiments/wandbdirectory by default. If you create a file at experiments/wandb.key containing your weights and biases key, then all the logs will be pushed to your wandb account.

If you find our paper or this repository useful, consider citing us at

@article{mate2024neural,
        author = {Mát{\'e}, Bálint and Fleuret, Fran{\c{c}}ois and Bereau, Tristan},
        title = {Neural Thermodynamic Integration: Free Energies from Energy-Based Diffusion Models},
        journal = {The Journal of Physical Chemistry Letters},
        volume={15},
        number = {45},
        pages = {11395-11404},
        year = {2024},
        doi = {10.1021/acs.jpclett.4c01958},
        note ={PMID: 39503734},
        URL = {https://doi.org/10.1021/acs.jpclett.4c01958},
        eprint = {2406.02313},
        pages={11395--11404},
}