Planet detection using Convolutional Neural Networks
- Robert Austin Benn
- Enrique Basañez Mercader
- Miguel Blanco Marcos
- Borja Sánchez Leirado
Analyze state-of-the-art research about planet detection using transit method (see https://en.wikipedia.org/wiki/Methods_of_detecting_exoplanets) and use advanced machine learning techniques to enble further discoveries.
For this project pipeline, different modules are used:
Module scr/com/saturdaysai/exonet/lightKurveApi/lightKurveApiCLient.py
Raw data will be obtained from Kepler records stored at MAST (Mikulski Archive for Space Telescopes) using Python library LightKurve to access their api. https://archive.stsci.edu/mast.html
Raw data obtained have informatioon about transit period, duration and centroid.
Using that information, data will be normalized and re-sampled to fixed length to generate one-dimensional tensors.
Notebook notebooks/model_definition.ipynb
This notebook uses tensors generated (and stored in /data folder) to define and optimize a CNN based model. Hyper-parameters present in this notebook are the result of iterative training.
Notebook notebooks/model_evaluation.ipynb
Loads model defined in previous step and validates it using validation sub-dataset, obtaining cross-entropy loss, accuracy, AUC, and precision vs recall curve.
Notebook notebooks/model_predictions.ipynb
Predictions will be made using Kepler Objects of Interest which have candidates of being exo-planets, ordering those by decreasing likelihood of having a light curve that evidences the presence of an exoplanet.
Article in Medium: https://medium.com/@miguel.blanco.marcos/exonet-an-ai-saturdays-project-a1bda907bdef
Inspired in: Shallue, Christopher J., and Andrew Vanderburg. “Identifying Exoplanets with Deep Learning: A Five-Planet Resonant Chain Around Kepler-80 and an Eighth Planet Around Kepler-90.” The Astronomical Journal 155.2 (2018): 94. Crossref. Web. (DOI: 10.3847/1538-3881/aa9e09)