PCFM.jl

Physics-Constrained Flow Matching: Sampling Generative Models with Hard Constraints

A Julia implementation for the NeurIPS 2025 Paper: Physics-Constrained Flow Matching: Sampling Generative Models with Hard Constraints

Note: This package is under active development. More features and capabilities will be added progressively.

Installation

using Pkg
Pkg.add(url = "https://github.com/utkarsh530/PCFM.jl")

Or in development mode:

Pkg.develop(path = "/path/to/PCFM")

Usage

For a complete working example, see examples/train_diffusion.jl.

To run the example:

julia --project examples/train_diffusion.jl

Citation

If you use this package, please cite:

@article{utkarsh2025physics,
  title={Physics-Constrained Flow Matching: Sampling Generative Models with Hard Constraints},
  author={Utkarsh, Utkarsh and Cai, Pengfei and Edelman, Alan and Gomez-Bombarelli, Rafael and Rackauckas, Christopher Vincent},
  journal={arXiv preprint arXiv:2506.04171},
  year={2025}
}

License

MIT License - see LICENSE file for details.