UMBRA

Deep learning detection of exoplanet transits in PSLS-simulated PLATO light curves

This project compares two methods for detecting planetary transits in simulated PLATO light curves: the Box Least Squares (BLS) algorithm and a ResNet-1D convolutional neural network. Both are tested on 1000 PSLS simulations. The CNN reaches 96.7% accuracy and 98.6% precision, compared with 91.3% accuracy for BLS. Both methods lose sensitivity for long-period planets that produce only one or two transits in the 267-day observing window.

Author: Zain Ul Abideen, Erasmus Mundus MASS programme GitHub repository: dominuszain/CompAstroBio

Results

Method	Accuracy	Confusion matrix
BLS	91.3%
ResNet-1D CNN	96.7%

The full analysis is in BLS/bls_transit_detection.ipynb and CNN/cnn_resnet_transit.ipynb. The report is in Docs/report.pdf.

Repository contents

Path	Contents
demo_bls.ipynb	Short BLS demo on the included sample
demo_cnn.ipynb	Short CNN demo on the included sample
full_bls_results.ipynb	BLS run on the full 1000-light-curve dataset
full_cnn_results.ipynb	CNN run on the full 1000-light-curve dataset
BLS/	BLS analysis notebook and notes
CNN/	ResNet-1D analysis notebook and notes
dataset/sample/	100 simulated light curves used by the demos (about 1.1 GB)
dataset/psls.yaml	PSLS configuration used to generate the data
dataset/script.sh	Script that produced the dataset
Task01/ to Task04/	Course task reports and the full metadata table
Docs/	Project report and figures
Tests/	Notebook execution and style checks

The sample directory holds 50 transit and 50 non-transit light curves, which is enough to run both demo notebooks end to end.

Installation

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

The CNN demo uses PyTorch and runs on a laptop CPU. No GPU is required.

Running the demos

Open demo_bls.ipynb or demo_cnn.ipynb in Jupyter and run all cells. If you move the dataset, update the DATA_DIR path at the top of each notebook.

Full dataset

The complete 1000-light-curve dataset (about 33 GB uncompressed, archived as an 11 GB zip) is available on Zenodo:

The data were generated with the PLATO Solar-like Light-curve Simulator (PSLS) using dataset/psls.yaml and dataset/script.sh. The full per-run metadata is in Task04/metadata.csv.

Tests

pip install pytest nbmake nbqa pylint
bash Tests/test_nbmake.sh
bash Tests/test_pylint.sh

Citation

If you use the code, paper, or dataset, please cite:

@misc{ulabideen_umbra_2026,
  author = {Ul Abideen, Zain},
  title  = {Deep Learning Detection of Exoplanet Transits in PSLS-Simulated PLATO Light Curves},
  year   = {2026},
  doi    = {10.13140/RG.2.2.31127.74407},
  url    = {https://doi.org/10.13140/RG.2.2.31127.74407}
}

@dataset{ulabideen_umbra_dataset_2026,
  author    = {Ul Abideen, Zain},
  title     = {PSLS-based computations for exoplanet transits},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19715574},
  url       = {https://doi.org/10.5281/zenodo.19715574}
}

License

The code is released under the MIT License. The simulated dataset and notebook outputs are released under CC BY 4.0.

Acknowledgements

Produced for the Computational Astrobiology course in the MASS programme (University of Belgrade, 2026).