Detecting Solar-like Oscillations in Red Giants with Deep Learning

TL;DR

This paper introduces a supervised deep learning approach to automatically detect solar-like oscillations in red giant stars and predict their frequency of maximum power, achieving high accuracy comparable to expert visual inspection.

Contribution

It presents a novel deep learning method trained on Kepler data to detect oscillations and estimate $ u_{ ext{max}}$, reducing reliance on manual analysis.

Findings

Detection accuracy of 98-99% on K2 data

Estimated $ u_{ ext{max}}$ uncertainty of about 5%

More robust $ u_{ ext{max}}$ estimates than classical methods

Abstract

Time-resolved photometry of tens of thousands of red giant stars from space missions like Kepler and K2 has created the need for automated asteroseismic analysis methods. The first and most fundamental step in such analysis, is to identify which stars show oscillations. It is critical that this step can be performed with no, or little, detection bias, particularly when performing subsequent ensemble analyses that aim to compare properties of observed stellar populations with those from galactic models. Yet, an efficient, automated solution to this initial detection step has still not been found, meaning that expert visual inspection of data from each star is required to obtain the highest level of detections. Hence, to mimic how an expert eye analyses the data, we use supervised deep learning to not only detect oscillations in red giants, but also predict the location of the frequency at maximum power, $\nu_{\mathrm{max}}$, by observing features in 2D images of power spectra. By training on Kepler data, we benchmark our deep learning classifier against K2 data that are given detections by the expert eye, achieving a detection accuracy of 98% on K2 Campaign 6 stars and a detection accuracy of 99% on K2 Campaign 3 stars. We further find that the estimated uncertainty of our deep learning-based $\nu_{\mathrm{max}}$ predictions is about 5%. This is comparable to human-level performance using visual inspection. When examining outliers we find that the deep learning results are more likely to provide robust $\nu_{\mathrm{max}}$ estimates than the classical model-fitting method.