Pushing the Limit of Asteroseismic Detection for Cool Dwarfs using TESS and Deep Learning

This paper presents a deep learning model trained on TESS light curves that achieves 99.8% accuracy in identifying solar-like oscillations in cool dwarfs, successfully narrowing down thousands of candidates to 24 promising stars to extend the detection frontier of asteroseismology for main-sequence and sub-giant stars.

The Gist

Imagine the universe is a giant, silent concert hall. For decades, astronomers have been trying to listen to the music of the stars. Some stars, like the giants, sing loud, deep notes that are easy to hear. But the smaller, cooler stars (like our Sun and even smaller "dwarf" stars) sing very quiet, high-pitched songs. These songs are called solar-like oscillations. They are the result of the star's surface churning like boiling water, creating tiny ripples that change the star's brightness ever so slightly.

The problem? These ripples are so faint that they get lost in the "static" of the universe, much like trying to hear a whisper in a hurricane.

Here is how the authors of this paper tackled that problem, explained simply:

1. The Challenge: Finding a Needle in a Haystack

Astronomers have a powerful space telescope called TESS that watches the sky, taking snapshots of stars every two minutes. It has collected data on thousands of stars. However, looking at this data by hand is like trying to find a specific needle in a haystack by looking at every single piece of hay one by one. The "needles" (the cool stars singing their quiet songs) are hidden among millions of other stars that are just noisy, spinning, or blinking for other reasons.

2. The Solution: Teaching a Digital Detective

Instead of looking at the raw video of the stars (the light curves), the authors decided to look at the sheet music (the periodogram). Think of a light curve as a recording of a song, and the periodogram as a graph showing which notes are being played.

• The "Signature" of a Star: A star singing a solar-like song has a very specific shape on this sheet music. It looks like a gentle hill (caused by the churning surface) with a distinct, hump-shaped peak sitting on top of it (the actual song).
• The AI Teacher: The authors built a computer program (a Convolutional Autoencoder) that acts like a student. They showed it thousands of examples of stars that do sing (the "good" students) and thousands of stars that don't (the "distracted" students).
• The Training: The computer learned to recognize the shape of that specific "hump" on the sheet music. It learned to ignore the static and the other types of noise.

3. The Results: A New List of Singers

Once the computer was trained, they let it loose on a massive list of 91,000 cool stars.

• The Filter: The computer acted like a super-efficient bouncer, instantly sorting the stars. It found 3,463 stars that looked like they might be singing.
• The Vetting: The human astronomers then took this list and did a final, careful check. They looked at the "sheet music" again to make sure the computer wasn't tricked by noise.
• The Final Cast: After all the checking, they found 24 stars that are very strong candidates for having these solar-like oscillations.

4. Why This Matters: Breaking the "Cool" Barrier

Most of the stars we know how to "hear" are big, hot, or evolved stars. The cool, small stars (like M-dwarfs and K-dwarfs) are usually too quiet to be heard with current technology.

• The Analogy: Imagine you have a microphone that can only hear loud singers. This paper is like teaching that microphone to hear the quietest, smallest singers in the room.
• The Discovery: Several of these 24 candidates are M-dwarfs (very small, cool stars). Finding them is a big deal because they are usually too faint to be studied this way. Some of these stars are so cool that they occupy a part of the "star map" that was previously only accessible if you used much more expensive and difficult tools (like measuring the star's wobble with giant telescopes).

5. The Caveat: "Potential" vs. "Confirmed"

The authors are careful to say they haven't definitively heard the song yet. They have found the candidates—the stars that look like they are singing based on the computer's analysis.

• The Next Step: To confirm these stars are actually singing, astronomers will need to do follow-up observations. They might need to use the telescope for longer periods or use different, more sensitive tools to catch the faint signal clearly.

Summary

In short, this paper is about using Artificial Intelligence to act as a super-powered filter. It taught a computer to recognize the unique "fingerprint" of quiet, cool stars. By doing this, they found a shortlist of 24 stars that might be singing songs we've never heard before, potentially opening a new chapter in our understanding of how small, cool stars are built and how they live.

Technical Summary

Technical Summary: Pushing the Limit of Asteroseismic Detection for Cool Dwarfs using TESS and Deep Learning

Problem Statement
Asteroseismology offers a powerful probe of stellar interiors by detecting solar-like oscillations, which are stochastically excited by near-surface convection. While thousands of such oscillators have been identified in evolved stars (red giants), the detection of these signals in main-sequence cool dwarfs (G, K, and M types) has been severely limited. These stars exhibit low oscillation amplitudes and high oscillation frequencies (often approaching the Nyquist limit of previous missions like Kepler and CoRoT). Consequently, only a handful of solar-like oscillators have been confirmed in the cool main sequence and sub-giant branches. The Transiting Exoplanet Survey Satellite (TESS) offers a unique opportunity due to its short 2-minute cadence (Nyquist limit ~4167 µHz) and sensitivity at red optical wavelengths, but searching large samples for these faint, stochastic signals using traditional parametric models is challenging due to the variability of signal morphology and the computational cost of manual verification.

Methodology
The authors developed a deep learning framework to automate the detection of solar-like oscillations in TESS 2-minute cadence light curves. The methodology proceeds through the following stages:

1. Data Representation: Instead of using raw time-series or phase-folded light curves, the study utilizes Lomb-Scargle periodograms. The authors argue that periodograms cleanly separate the broadband granulation background from the oscillation power excess. To optimize for neural network processing, these periodograms are binned logarithmically and resampled to a fixed length of 4096 points.
2. Training Data Construction:
   • Positive Class (SOLR): A sample of 4,177 manually verified solar-like oscillators from E. Hatt et al. (2023), derived from TESS 120-s and 20-s cadence data.
   • Negative Class (NonSOLR): A sample of 82,204 stars representing other variability types (pulsators, rotators, eclipsing binaries) and non-variables, sourced from L. A. Balona (2023).
   • Preprocessing: Light curves were cleaned of outliers (>5σ), and periodograms were normalized using MinMaxScaler.
3. Network Architecture: The core model is a one-dimensional convolutional autoencoder coupled with a classification head.
   • Autoencoder: Designed to learn a compressed latent representation (embedding) of the periodogram morphology. The architecture was deepened (14 convolutional layers total) compared to previous works to handle the 4096-point input. It utilizes BatchNormalization, Leaky ReLU activations, and MaxPooling for dimensionality reduction. The latent space dimension was optimized to 128 to balance reconstruction fidelity (specifically capturing the oscillation "hump") and classification performance.
   • Classifier: A Multi-Layer Perceptron (MLP) attached to the latent space to distinguish between SOLR and NonSOLR classes.
   • Training Strategy: The network was trained sequentially: first, the autoencoder was trained to minimize Mean Squared Error (MSE) reconstruction loss; subsequently, the classifier was trained on the frozen latent features using Binary Cross-Entropy with Logits (BCEWithLogits).
4. Candidate Vetting: The trained network was applied to the Asteroseismic Target List (ATL), selecting stars cooler than G0 ($T_{eff} \le 5930$ K). Candidates with a predicted probability $>0.5$ underwent manual vetting using the pySYD package and echelle diagnostics. Criteria included:
   • Consistency between observed $\nu_{max}$ and ATL predictions.
   • Presence of periodicity in the autocorrelation function (ACF).
   • Absence of background contamination.
   • Presence of vertical ridges in the échelle diagram.

Key Results

• Model Performance: The neural network achieved a classification accuracy of 99.8% on the independent test set, with a Precision of 0.945, Recall of 0.998, and an F1 Score of 0.971.
• Candidate Identification: From an initial sample of ~81,642 stars, the model identified 3,463 potential solar-like oscillators (probability > 50%).
• Final Candidate List: After rigorous vetting, the authors identified 24 candidate stars (23 from 2-minute cadence data and 1 from 20-second cadence data) that satisfy all diagnostic criteria.
  • Cool Dwarfs: Notable candidates include two M-dwarfs (TIC 240764948 and TIC 406430692) and several K-dwarfs (e.g., TIC 84332248, TIC 34115230) located in regions of the color-magnitude diagram previously inaccessible to TESS asteroseismology.
  • Sub-giants: Candidates such as TIC 83489517 and TIC 402363298 occupy regions slightly above the main sequence where no prior TESS detections existed.
• 20-Second Cadence: A supplementary search using 20-second cadence data for a subset of targets yielded one new candidate (TIC 97036607) and significantly improved the signal-to-noise ratio (SNR) and diagnostic clarity for three existing candidates.

Significance and Claims
The paper claims that this work represents the first deep-learning study specifically targeting solar-like oscillations in cool main-sequence and sub-giant dwarfs using TESS data. The significance lies in:

1. Extending the Detection Frontier: The identified candidates occupy regions of the color-magnitude diagram (particularly cool K and M dwarfs) that are typically only accessible via resource-intensive radial velocity observations.
2. Methodological Validation: The study demonstrates that convolutional autoencoders can effectively learn the morphological signatures of solar-like oscillations (granulation slope + Gaussian power excess) in periodograms, outperforming traditional parametric approaches for initial screening.
3. Foundation for Future Work: The authors explicitly state they do not claim definitive detections but rather provide a validated list of candidates. These targets serve as a foundation for follow-up observations (higher cadence photometry and radial velocity) to confirm oscillations, which would ultimately improve understanding of stellar structure and evolution across the lower main sequence.

The paper concludes that deep learning techniques offer a robust, scalable method for studying stellar light curves, potentially expanding the sample of cool-dwarf solar-like oscillators and strengthening the evidence for using such techniques in time-domain astronomy.