GANASTRO: a data-driven approach for extracting exoplanetary atmospheric features

Official code repository for the paper "A data-driven approach for extracting exoplanetary atmospheric features" - M. Giordano Orsini et al., 2025

• Paper figures are stored in the folder paper/images and in experiments/inpainting subfolders.
• Experimental results are extracted from text files in experiments/comparison.
• The file make_plots.py provides the code for reproducing some plots shown in the paper.

Reproducing the experiments

We provide instructions in order to reproduce the experimental results of the proposed work.

Installation

conda env create -f environment.yml
conda activate ganastro

Data manipulation

Prerequisite: a set of real nights extracted from some data archive.

Launch augment.py for generating a synthetic dataset for model training.

Training the DCGAN model

To train a DCGAN model, first you need to provide a configuration file. The json file dcgan.json contains the model specifications actually used for obtaining the experimental results proposed in the paper.

python train.py --config dcgan.json

In a time-limited session, you can resume the model training for a given checkpoint by using:

python train.py --resume <path-to-checkpoint> --config <path-to-config>

N.B.: To get the weights of the pretrained model used for the experiments in the paper, please contact me at [email protected]

Please put the downloaded file in pretrained/.

Semantic image inpainting using DCGANs

Once trained, the generator $G$ approximates the distribution of transit-free nights, given the input real nights.

To perform semantic image inpainting, first you need to prepare a configuration file for the inpainting settings. The json file inpainting.json provides the inpainting configuration used in our work.

python inpaint.py --resume pretrained/0920_170346/checkpoint-epoch250.pth --config inpainting.json

Running an experimental comparison with PCA

The following schema describes the comparison process between DCGAN and PCA-based detrending algorithm. More details in the paper.

We perform the PCA decomposition by computing the optimal number of principal components retaining the 90% of the total variance. However, this hyperparameter can be specified by setting the option --energy_threshold.

To perform the comparison, launch:

python comparison.py --energy_threshold 90

Citation

If you found this work useful for your research, please consider citing the corresponding paper:

@article{giordanoorsini_detrending_2025,
title = {A data-driven approach for extracting exoplanetary atmospheric features},
journal = {Astronomy and Computing},
volume = {52},
pages = {100964},
year = {2025},
issn = {2213-1337},
doi = {https://doi.org/10.1016/j.ascom.2025.100964},
url = {https://www.sciencedirect.com/science/article/pii/S221313372500037X},
author = {Massimiliano {Giordano Orsini} and Alessio Ferone and Laura Inno and Paolo Giacobbe and Antonio Maratea and Angelo Ciaramella and Aldo Stefano Bonomo and Alessandra Rotundi},
keywords = {Exoplanetary atmospheres, High-resolution spectroscopy, Deep learning, Detrending},
}