Split-Flows: Measure Transport and Information Loss Across Molecular Resolutions

Overview

Split-flows provide a probabilistic bridge between molecular resolutions, enabling conditional backmapping and direct measurement of the configuration-dependent (local) information loss.

Installation

Clone the repository and navigate to the project directory:

git clone git@github.com:hummerichsander/split-flows.git
cd split-flows

To install the project dependencies use uv (if you have not installed uv yet, check out the uv documentation) and run:

uv sync

Usage

Training

Model training can be done using the hydrantic package, which bundles pytorch-lightning, hydra, and pydantic for model specification and training.

To let hydrantic know about the location of the configuration files you can set the environment variable HYDRANTIC_CONFIG_PATH to the path of the config directory.

Training a model can be done using the hydrantic command line interface. To train a model for alanine dipeptide (ala2.yml) run:

python -m hydrantic.cli.fit --config-name ala2

Model loading

Weights and hyperparameters are stored as checkpoints (.ckpt files). To instantiate a model from a checkpoint, use the load_from_checkpoint method of the Model class:

from split_flows.models import SplitFlow

model = SplitFlow.load_from_checkpoint(<checkpoint path>)

Citation

If you use split-flows in your research, please cite:

@misc{hummerich2025splitflowsmeasuretransportinformation,
      title={Split-Flows: Measure Transport and Information Loss Across Molecular Resolutions},
      author={Sander Hummerich and Tristan Bereau and Ullrich Köthe},
      year={2025},
      eprint={2511.01464},
      archivePrefix={arXiv},
      primaryClass={physics.chem-ph},
      url={https://arxiv.org/abs/2511.01464},
}