Identify exoplanet candidates with residual networks and fully connected neural networks

Abstract

The work is about identifying exoplanet candidates using deep learning models including two residual networks and one fully connected neural network. Finding these objects manually is a very labor intensive task. Therefore using reliable algorithms to manage the data is necessary. Deep learning can be helpful because it suits well for very large input data. In general, having more data generally makes deep learning models perform better. We use Kepler Space Telescope data to detect planet candidates by using convolutional neural network models. We apply the Kepler Q1-Q17 (DR24) table as our training and test sets. The model takes two phase-folded light curves and some parameters of each transit-like signal and then outputs whether the signal represents a planet candidate (PC), a non-transiting phenomena (NTP) or a false positive (FP). In the current model, we feed 45 features into a fully connected neural network, such as transit durations and depth of signals. At this stage, the models achieve AUC (Area Under ROC Curve) and accuracy of about 98:4% and 96:7% respectively for the test set. The accuracy for the training set can be over 99%, which means that the model can easily over- t the training data. The most straightforward way to the problem is to use more data to train the model. Therefore, we plan to train it with more simulated data later in order to increase the AUC, accuracy, precision and recall of the predictions.