This repository provides an efficient implementation in jax of a score matching and physics informed neural network-based algorithm for solving the stationary Fokker-Planck equation in high dimension.
The implementation is built on Google's jax package for accelerated linear algebra and DeepMind's haiku package for neural networks. Both can be installed by following the guidelines at the linked repositories.
Routines common to all implemented simulations can be found in py/common, including implementations of the various neural networks used, systems studied, and loss functions used.
Simulation code to launch learning experiments can be found in py/launchers.
Code for generating datasets can be found in py/dataset_gen.
Code for visualizing the output of simulations and for producing the publication figures can be found in notebooks.
Slurm sbatch scripts used to launch the experiments in the paper can be found under slurm_scripts.
Experiment tracking is implemented in Weights and Biases. You will need to input a project title in the corresponding simulation launcher in the call to wandb.init.
If you found this repository useful, please cite:
[1] N. M. Boffi and Eric Vanden-Eijnden. “Deep learning probability flows and entropy production rates in active matter", arXiv: 2309.12991.
@misc{boffi2023deep,
title={Deep learning probability flows and entropy production rates in active matter},
author={Nicholas M. Boffi and Eric Vanden-Eijnden},
year={2023},
eprint={2309.12991},
archivePrefix={arXiv},
primaryClass={cond-mat.stat-mech}
}