Description
The discovery of exoplanets is vital to understanding planetary systems beyond our solar system, yet less than 0.1% of galactic stars have been surveyed, according to NASA. Light curves, which track a star’s brightness, or flux, over time, can reveal periodic dimming indicative of potential exoplanet transits. However, this dimming can result from various other astrophysical phenomena, meaning false positives are common when evaluating flux-over-time data for exoplanet candidacy. To more accurately identify exoplanet candidates using light curves from the TESS mission, a machine-learning pipeline was developed. Astrophysical data for 38,047 confirmed and potential exoplanets were sourced from NASA’s Exoplanet Archive. A hybrid deep learning model was created, featuring convolutional, long short-term memory, dropout, max pooling, and dense layers. The convolutional layers were used to evaluate local patterns and variations in flux, such as sudden drops in brightness or gradual changes over time, while the long short-term memory layer was responsible for analyzing long-range temporal patterns throughout the light curve. Dropout, max pooling, and dense layers helped prevent overfitting, reduce dimensionality, and refine the final classification, respectively. The model was trained on this data to learn patterns that distinguish exoplanetary signals from other astrophysical phenomena, reaching 97% accuracy. A total of 12,195 light curves from NASA's Transiting Exoplanet Survey Satellite mission were then evaluated using the model. The model’s evaluation resulted in the identification of 256 new exoplanet candidates. However, further observations and validation are necessary to confirm their planetary status.
