Exoplanet Detection: Analyzing Kepler Telescope Time-Series Data with Convolutional Neural Networks

ExoPlanet-Detection is an open-source project that implements automatic exoplanet detection using Convolutional Neural Networks (CNN) based on light curve data from the Kepler Telescope. This project applies deep learning technology to astrophysics, providing efficient automated tools for analyzing massive astronomical data and supporting exoplanet research.

Since the discovery of the first exoplanet in 1995, the transit method has become the main detection method for telescopes like Kepler. However, manual analysis of the massive light curve data generated by Kepler (tens of thousands of stars, years of observations) is difficult to complete, so machine learning technology is urgently needed to solve this problem.

Kepler light curves record changes in stellar brightness. Transit events cause periodic brightness dips (usually less than 1%), but the data has issues like noise and outliers. CNN can automatically learn transit signal features (such as U-shaped/V-shaped dips), which is more robust than manual feature engineering and suitable for large-scale data processing.

The project includes modules such as data acquisition (01_descarga.py), preprocessing (02_preprocesar.py), model training (03_entrenamiento.py, ExoNet.py), experimental analysis (multiple Notebooks like Shallue_model.ipynb), and hyperparameter optimization (Optuna framework), covering the complete MLOps workflow.

1. Draws on the groundbreaking work of Google researchers in 2018; 2. Explores models combining wavelet transform and time-frequency analysis; 3. Follows good machine learning engineering practices, such as data separation, version control, and hyperparameter tracking.

1. Assists astronomers in quickly screening high-confidence candidate planets; 2. May discover weak-signal planets missed by traditional methods (e.g., Kepler-90i); 3. Experience can be applied to data analysis of next-generation telescopes like TESS.

1. Data class imbalance (positive samples are far fewer than negative samples); 2. False positive problem (non-planetary phenomena easily produce similar transit signals); 3. Machine learning results require manual verification by professional astronomers.

This project demonstrates the potential of machine learning in astronomical research and provides an efficient tool for exoplanet detection. With the commissioning of facilities like JWST, machine learning will play a greater role in planetary atmosphere characterization and habitable world search, and this open-source project provides an entry reference for developers and researchers.