Exoplanet detection with Genesis

Abstract

Convolutional Neural Networks (CNNs) have shown to offer a consistent and reliable foundation for the automatic detection of potential exoplanets. CNNs rely on an abundance of parameters (overparameterization) to achieve their impressive detection performances. Astronet was one of the first CNNs for exoplanet detection. It takes as input folded lightcurves in two views: a local view (the transit) and a global view (the entire orbital period including the transit). A more recent CNN called Exonet-XS improved on Astronet’s performance while having considerably less parameters, thereby reducing the risk of overfitting. Exonet-XS also uses two views as input. In this paper, we propose Genesis, an even more simplified CNN for exoplanet detection from folded lightcurves using only one view. In addition, we propose to use a more reliable validation procedure that is custom in CNN-based exoplanet detection studies: the Monte Carlo Cross-Validation (MCCV) procedure. We show that the use of MCCV improves the reliability of the estimation of the detection performance by providing a (discretized) probability distribution, rather than a point estimate. Using MCCV we show that Astronet with only one view performs on a par with the original two-view version. More importantly, our fair comparative evaluation (without stellar parameters and centroids) reveals that Genesis outperforms Exonet-XS and Astronet. We conclude by stating that existing exoplanet detection CNNs are too complex for the task at hand and that future evaluations of performances should use MCCV or similar validation procedures.

Size does matter: Exoplanet detection with a sparse convolutional neural network

“…The second row reveals the main architectural difference between Genesis and both Exonet architectures. Where the latter have two input views, one processing a global view and one a local view of the folded light curve, Genesis relies on a single input view (Visser et al, 2022). The use of a single view leads to a reduced number of convolution layers (4 instead of 14 and 5 for Exonet and Exonet-XS, respectively).…”

“…For our experiments we made use of the data set used by Ansdell et al (2018) that followed the procedure of Shallue and Vanderburg (2018) (see Visser et al, 2022 for further details). The data set is based on the Q1-Q17 Data Release 24 (DR24) light curves from the Mikulski Archive for Space Telescopes (Thompson et al, 2015) and consists of 15,084 light curves, i.e., long cadence brightness values obtained by fixed-aperture photometry by summing over the pixel values within the aperture (Twicken et al, 2010).…”

“…Such a partitioning results in point estimates of the accuracy and AUC, which hampers a proper comparative evaluation. Hence, as in our previous work (Visser et al, 2022), we employ Monte Carlo Cross Validation (MCCV). Following this procedure, training and evaluation of an architecture is repeated n times (in our case, n = 20).…”

“…All experiments were performed with the same settings discussed in Visser et al (2022). Each model instance is trained for a maximum of 125 epochs, with a batch size of 1024.…”

“…In previous work, we presented Genesis (Visser et al, 2022), an even sparser architecture than Exonet-XS. We performed a careful comparative evaluation, using a cross-validation procedure, to obtain performance distributions rather than point estimates.…”