Transiting Exoplanet Hunting Using Convolutional Neural Networks

Abstract

The human race has spent and invested considerable resources to understand the processes of your solar system. However, since the discovery of exoplanets we have the means to determine whether or not we actually understand these processes. The most compelling reason to find exoplanets is that it opens the door for us to look for other habitable planets as well as understand our own solar system better. For years, scientists have been utilizing data from NASA’s Kepler Space Telescope to look for and identify thousands of transiting exoplanets. Thanks to new and better telescopes, astronomical data is rapidly increasing. Traditional human judgment-based prediction and classification methods are inefficient and vulnerable to vary depending on the expert doing the study. The widely used methodology for exoplanet discovery, the Box-fitting Least Squares technique (BLS), for example, creates a large number of false positives that must be manually checked in the event of noisy data. As a result, an automated and unbiased approach for detecting exoplanets while removing false-positive signals imitating transiting planet signals is required. A new convolutional neural network-based mechanism for finding exoplanets is introduced using the transit technique. Since the dataset is large and highly imbalanced, SMOTE is used to resample the data, while the exponential decay approach along with dropout and early stopping techniques are used to reduce model overfitting. In addition, the model employs the Grid-SearchCV approach to fine-tune hyper-parameters. Finally, for a robust and full model, the model is evaluated using k fold cross-validation. Performance criteria such as accuracy, precision, recall, f1 score, sensitivity, and specificity are used in the study. After analyzing the data, the research concluded that the convolutional neural network produced a maximum accuracy of 99.6% on the testing data.