Neural networks as exoplanet fishing nets.

NIETO, L. A.; SEGURA, E.; DÍAZ, R. F.

Workshop; Fourth Workshop on Extreme Precision Radial Velocities; 2019

The number of applications of neural networks (NN) in science is increasing every day. Exoplanet research is not an exception, with many machine learning algorithms used to detect and / or classify transiting planet candidates, or to characterise exoplanet atmospheres. But the use of NN analyse Extreme Precision Radial Velocity (EPRV) data is still limited. Here I present our first attempts at building a NN that detects low-amplitude periodic signals in time series having the typical characteristics of EPRV data. We used both convolutional and recursive NN, and addressed the problem first as a classification and then as a regression one. We trained the NN with synthetic EPRV time series, but this gives mediocre results. On the other hand, using periodograms as input, much better and promising results are reached. In particular, NN perform substantially better than a virtual researcher that explored the data automatically using the traditional method of finding the highest peak in the periodogram, removing a signal at this period, and repeating until no peak power in the periodogram has p-value below a given level.