Flow Matching for Atmospheric Retrievals

This repository contains the code for the research paper:

T. D. Gebhard, J. Wildberger, M. Dax, A. Kofler, D. Angerhausen, S. P. Quanz, B. Schölkopf (2024). "Flow Matching for Atmospheric Retrieval of Exoplanets: Where Reliability meets Adaptive Noise Levels." Accepted for publication in Astronomy & Astrophysics. Available on arXiv.

---

🚀 Quickstart

Installation should generally work by checking out this repository and running pip install on it:

git clone git@github.com:timothygebhard/fm4ar.git ;
cd fm4ar ;
pip install -e .

For developer mode (e.g., unit tests, linters, ...), replace the last line with:

pip install -e ".[dev]"

🏕 Setting up the environment

The code in here relies on some environmental variables that you need to set:

export FM4AR_DATASETS_DIR=/path/to/datasets ;
export FM4AR_EXPERIMENTS_DIR=/path/to/experiments ;

You might want to add these lines to your .bashrc or .zshrc file.

Generally, these folders can be subfolders of this repository; however, there may exists scenarios where this is not desirable (e.g., on a cluster).

🐭 Tests

This repository comes with a rather extensive set of unit tests (based on pytest). After installing ml4ptp with the [develop] option, the tests can be run as:

pytest tests

You can also use these tests to ensure that the code is still working when you update the dependencies in pyproject.toml.

📜 Citation

If you find this code useful, please consider citing our paper:

@article{Gebhard_2024,
  author     = {Gebhard, Timothy D. and Wildberger, Jonas and Dax, Maximilian and Angerhausen, Daniel and Quanz, Sascha P. and Schölkopf, Bernhard},
  title      = {Flow Matching for Atmospheric Retrieval of Exoplanets: Where Reliability meets Adaptive Noise Levels},
  year       = 2024,
  journal    = {Astronomy \& Astrophysics},
  eprint     = {2410.21477},
  eprinttype = {arXiv},
  addendum   = {(Accepted)},
}

⚖️ License and copyright

The code in this repository was written by Timothy Gebhard, with contributions from Jonas Wildberger and Maximilian Dax, and is owned by the Max Planck Society. We release it under a BSD-3 Clause License; see LICENSE for more details.