Um estudo sobre a detecção de exoplanetas com inteligência artificial

Abstract

Space missions like CoRoT, Kepler and TESS are essential in astrophysics for the search of exoplanets. Thousands of planets have already been detected with the improvement of search techniques. However, most exoplanets were discovered by visual inspection of the light curves obtained with the CoRoT and Kepler missions. These processes are subject to human error and are impractical when a large volume of data is involved. The future PLATO and James Webb Space Telescope space missions will bring even greater volumes, which presents a serious challenge to the research methodology that was being done so far. Based on this, new methods are needed to treat and analyze the photometric light curves coming from these missions. The use of artificial intelligence has been gaining more and more relevance in data science focused on astrophysical research and, particularly, it shows itself as a promising tool for the automatic detection of exoplanets. This work aims to apply Machine Learning for the detection of exoplanets based on the data generated by the Kepler mission. To this end, computer models such as Random Forest and Neural Networks were tested in the implementation of an algorithm that will be used to explore the light curves of any mission. Regarding the present research, more than 15 thousand light curves were divided into three sets: one for training the models, another for their validation and another for testing. It is previously known which light curves have confirmed planets or not. Preliminary results show that our best implemented model was the Neural Network, with an accuracy of 90% in the classification of the curves in the test set. It was also able to detect two planets confirmed in additional noisy data. This indicates that new cases of interest can be discovered by implementing our algorithm on an unexplored dataset. This is an important perspective of our work when data from the PLATO and James Webb missions become available.