Abstract
Abstract The precise derivation of transit depths from transit light curves is a key component for measuring exoplanet transit spectra, and henceforth for the study of exoplanet atmospheres. However, it is still deeply affected by various kinds of systematic errors and noise. In this paper we propose a new detrending method by reconstructing the stellar flux baseline during transit time. We train a probabilistic long short-term memory (LSTM) network to predict the next data point of the light curve during the out-of-transit, and use this model to reconstruct a transit-free light curve—i.e., including only the systematics—during the in-transit. By making no assumption about the instrument, and using only the transit ephemeris, this provides a general way to correct the systematics and perform a subsequent transit fit. The name of the proposed model is TLCD-LSTM, standing for transit light-curve detrending-LSTM. Here we present the first results on data from six transit observations of HD 189733b with the IRAC camera on board the Spitzer Space Telescope, and discuss some of its possible further applications.
Highlights
Since the first exoplanet atmosphere observation 20 years ago (Charbonneau et al 2000), more than 3000 transiting extrasolar planets have been discovered
We presented a deep learning model suitable for interpolating time series, and showed how it can be used to predict the variability of stellar light curves for subsequent transit fits
The presented method is similar to the Gaussian processes (GPs) approach (Rasmussen & Williams 2005; Gibson et al 2012) in that they both construct highly nonlinear models, avoid explicit physical modeling of the systematics, and provide probabilistic predictions
Summary
Since the first exoplanet atmosphere observation 20 years ago (Charbonneau et al 2000), more than 3000 transiting extrasolar planets have been discovered. The main instrumental systematics trend observed both with the Hubble Space Telescope WFC3 and the Spitzer IRAC cameras are the so-called ramp effect (Knutson et al 2007), hypothesized to be due to the charge trapping in the detector (Agol et al 2010), and intra-pixel and inter-pixel variations, which are correlated with the position of the source on the detector that shows variations in quantum efficiency across different pixels.1 Footprints of these entangled variability sources can be found in additional instrumental data collected besides the detector raw flux. We make use of an LSTM neural network (Hochreiter & Schmidhuber 1997) to interpolate the flux of a raw light curve during the transit, given additional time-series data coming from the PSF centroid.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
