Modular Machine Learning Framework for Automated Asteroseismic Characterization of TESS Red Giants

Ghanghas, Nipun; Palakkatharappil, Dinil Bose; Hanasoge, Shravan; García, Rafael A.; Liagre, Bastien; Mathur, Savita
Referencia bibliográfica

TASC10/KASC17 - 10th TESS/17th Kepler Asteroseismic Science Consortium Workshop

Fecha de publicación:
10
2026
Número de autores
6
Número de autores del IAC
1
Número de citas
0
Número de citas referidas
0
Descripción
Asteroseismology probes the internal structure of stars and yields precise constraints on their mass, radius, and age. The TESS mission, with its near-full-sky coverage, is expected to yield over 300,000 oscillating red giants, many with multi-sector observations. However, extracting high-precision seismic parameters at this scale requires robust, automated techniques capable of handling varying signal-to-noise ratios (SNR). We present a modular machine learning framework designed for the automated inference of the frequency of maximum power (νmax) and the large frequency separation (∆ν). Our approach utilizes a multi-stage architecture: first, a 1D Convolutional Neural Network (CNN) infers νmax directly from the raw power spectral density (PSD) without manual preprocessing. This νmax value is then used to dynamically crop the PSD for a second specialized network. To achieve high precision in ∆ν, this second network integrates both the flattened PSD and the Power Spectrum of Power Spectrum (PSPS). By leveraging the PSPS, which peaks at harmonics of ∆ν, our model can infer ∆ν very precisely and maintains robustness even for low-SNR targets where a direct PSD-based inference could fail. We demonstrate the efficacy of this pipeline on a sample of approximately 20,000 TESS red giants with 10 or more sectors of data, showing strong agreement with auto-correlation based inferences for high-SNR cases while providing superior reliability in the low-SNR regime. Our current efforts are focused on extending this high-precision PSPS-based inference to stars with as few as three sectors of TESS data. Furthermore, we are developing a specialized module to extract mixed-mode parameters, specifically the period separation of gravity-modes (∆Π) and the coupling factor (q). This module targets high-confidence, high-SNR stars identified in earlier stages of the pipeline for mixed-mode characterization where signal quality permits. Overall, this framework establishes a scalable, automated solution for analyzing massive datasets from TESS and future large-scale surveys such as Roman and PLATO.