2026 workshop program
Machine learning for exoplanet detection using the radial velocity method
The work
Abstract
Only 30 years ago, the first exoplanet orbiting a Sun-like star was discovered. To date, more than 5900 have been detected, and the number continues to grow rapidly thanks to technological advances.
The radial velocity (RV) method has proven to be one of the most successful and promising techniques for detecting planets through the motions they induce on their host stars. While instrumental improvements have enabled the measurement of increasingly smaller velocity variations, stellar activity and irregular sampling can make the detection of planetary signals more difficult and lead to false positives. For this reason, machine learning techniques have recently begun to be explored to address this challenge.
In this thesis, we develop a convolutional neural network with an attention layer to detect planetary signals in Sun-like stars, using simulated RV measurements generated over observation calendars representative of exoplanet searches. The network achieves 54% fewer false positives than the traditional null-hypothesis-based approach, without increasing the number of false negatives. This improvement is mainly concentrated in low amplitude signals, associated with low mass planets. In addition, the attention layer weights were analyzed to identify which regions of the input the model prioritizes during classification, revealing a correlation between these weights and the network’s predictions.
The method was further evaluated on 159 real signals from stars with at least one confirmed planet and achieved correct classifications in most cases. These data were also used to perform fine-tuning on the network, enhancing its detection capability on real observations. Overall, these results highlight the potential of neural networks as a promising tool for the detection of planetary signals in radial velocity measurements.