🔭 astrophysics

ABYSS. IV. Identifying signatures of stellar youth in APOGEE spectra

This paper presents a convolutional neural network classifier that effectively identifies stellar youth (under 40 Myr) in APOGEE and BOSS spectra by detecting rotational broadening and spotted photosphere signatures, thereby enabling robust separation of pre-main-sequence stars from evolved sources to reduce contamination in young stellar object candidates.

Original authors: Valentina Bonilla Villalobos, Marina Kounkel, Joseph Mullen, Eleonora Zari, Alexandre Roman-Lopes, Keivan Stassun, Ricardo López-Valdivia, Jinyoung S. Kim, Mojgan Aghakhanloo, Facundo Pérez Paolino, J

Published 2026-08-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Valentina Bonilla Villalobos, Marina Kounkel, Joseph Mullen, Eleonora Zari, Alexandre Roman-Lopes, Keivan Stassun, Ricardo López-Valdivia, Jinyoung S. Kim, Mojgan Aghakhanloo, Facundo Pérez Paolino, Jonathan C. Tan, Jesús Hernández

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Young stars are the universe's newborns, still glowing with the heat of their formation and surrounded by the dusty disks from which planets may eventually emerge. To astronomers, finding these infants among the billions of older, more settled stars in our galaxy is a difficult task. Young stars often look deceptively similar to their older cousins, sharing the same colors and brightness levels that usually help scientists sort them out. While some young stars show clear signs of youth, such as specific chemical fingerprints or rapid spinning, these clues are often hidden or hard to detect when looking through the thick clouds of dust that obscure the galactic plane. This makes it challenging to build a complete picture of how stars are born and how they evolve in their earliest years.

A team of researchers has now developed a new way to spot these stellar infants using a type of computer program known as a neural network. This tool was trained to analyze the light spectra of stars—essentially the chemical fingerprints left behind when starlight is split into its component colors—collected by the Apache Point Observatory Galactic Evolution Experiment, or APOGEE. This massive survey has already gathered spectra for hundreds of thousands of stars, but distinguishing the young ones from the old ones has remained a problem, especially for stars that are not extremely cool. The researchers built a classifier that scans the entire spectrum of a star at once, looking for subtle patterns that human eyes or traditional methods might miss.

The system proved remarkably effective, particularly at identifying young stars that are cooler, such as the red and orange dwarfs that make up the majority of stars in our galaxy. It successfully separated these pre-main-sequence stars from older field stars, even when the two groups had nearly identical temperatures and surface gravities. The model achieved its best results with these cooler stars, correctly identifying about seventy percent of the known young stars in its training set while maintaining a high level of accuracy in avoiding false alarms. When applied to new data that the computer had never seen before, it uncovered thousands of additional candidates, many of which were clustered around known regions where stars are currently forming, such as the Orion constellation.

To understand how the computer made these decisions, the researchers examined the specific features in the starlight that triggered the identification. They found that the program relied heavily on two main clues. The first was the broadening of spectral lines, which occurs when a star spins rapidly. Young stars tend to spin much faster than older stars, which have slowed down over billions of years, and this rapid rotation smears out the sharp lines in their spectra. The second clue was more subtle and related to the surface of the star itself. Young stars are often covered in large, dark spots caused by intense magnetic activity, similar to sunspots but on a much grander scale. Because these spots are cooler than the rest of the star's surface, they alter the shape of the spectral lines in a way that is distinct from older, calmer stars.

By combining these two signals—the blur of rapid rotation and the specific distortions caused by magnetic spots—the classifier could distinguish a young star from an older one even when they looked otherwise identical. The researchers identified dozens of specific wavelengths where these differences were most pronounced, creating a new list of potential youth indicators for the near-infrared part of the spectrum. While the exact identity of every single line remains a subject for further study, the pattern is clear: the light from a young star carries a unique signature of its youth that persists even as it approaches the main sequence.

This work provides a powerful new tool for astronomers to clean up their catalogs of young star candidates. Previously, many surveys relied on photometric methods that could be easily confused by reddened old stars or binary systems. The new classifier offers a more robust way to filter out these imposters, allowing researchers to focus their efforts on the most promising candidates. As the survey continues to gather data, this method will help build a more accurate and complete census of the young stellar population in our galaxy, offering a clearer view of the chaotic and vibrant early stages of stellar life.

Technical Summary: ABYSS. IV. Identifying signatures of stellar youth in APOGEE spectra

Problem Statement
Identifying young pre-main-sequence (PMS) stars within the field of evolved stars is a non-trivial challenge. While young stars possess distinct physical properties (e.g., larger radii, higher luminosities), they often overlap with binary-sequence stars and reddened red giants in the Hertzsprung-Russell diagram, particularly in terms of effective temperature ( $T_{\text{eff}}$ ) and surface gravity ( $\log g$ ). Traditional spectroscopic diagnostics for youth, such as the Li I 6707 Å line and H $\alpha$ emission, are prominent in the optical regime but are often obscured by extinction in embedded star-forming regions. Conversely, the Apache Point Observatory Galactic Evolution Experiment (APOGEE) provides high-resolution near-infrared (H-band) spectra for vast numbers of stars, including those in dusty environments. However, robust spectroscopic diagnostics for youth in the near-IR are less established than their optical counterparts, creating a need for methods to confirm the youth of candidates observed by APOGEE without relying solely on optical features.

Methodology
The authors developed a Convolutional Neural Network (CNN) classifier designed to operate on the full spectral range of APOGEE spectra to identify stellar youth ( $<40$ Myr).

Data Sources: The study utilizes data from the SDSS-V APOGEE & BOSS Young Star Survey (ABYSS). The training set was constructed from sources targeted by ABYSS (labeled "likely young stars") and the remaining APOGEE sources (labeled "field stars"). To mitigate cross-contamination inherent in these labels, the authors performed an iterative re-labeling process. This involved excluding obvious evolved stars based on $T_{\text{eff}}$ and $\log g$ (using MIST isochrones to exclude stars $>40$ Myr) and manually identifying spatial and kinematic overdensities in known star-forming regions (e.g., Orion, Taurus) to reincorporate likely YSOs that had been initially excluded.
Model Architecture: The CNN architecture, adapted from previous work on SDSS spectra, processes the entire spectrum as input. It includes an initial 1D convolution, ELU activation, and Max Pooling, followed by four Residual Network blocks (each with two 1D convolutional layers, batch normalization, and ELU activation). The data is then passed through adaptive average pooling and six fully connected layers with ReLU activation.
Training Strategy: The model was trained using a sparse categorical cross-entropy loss function. The authors intentionally preserved the intrinsic class imbalance (young stars are a small fraction of the total sample) to avoid biasing the model toward false positives, prioritizing a conservative assessment that relies on unique spectral features.
Feature Analysis: To understand the physical basis of the classifier's decisions, the authors constructed empirical templates by stacking spectra of young and field stars within identical bins of $T_{\text{eff}}$ , $\log g$ , and $[\text{Fe/H}]$ . They compared these templates to isolate spectral differences, specifically testing the impact of rotational broadening by convolving field star templates with rotational kernels.

Key Contributions and Results

Classifier Performance: The CNN successfully identifies young stars across a wide range of spectral types, with the strongest performance observed in M and K dwarfs.
- Recall: The model achieves $\sim70\%$ recall on the labeled training set, with recall decreasing as a function of $T_{\text{eff}}$ (nearly 100% for M dwarfs, dropping to $\sim0\%$ for O-type stars).
- Precision: For sources identified as young but not originally labeled as such ("false positives"), the precision is $\sim90\%$ . Many of these recovered sources exhibit spatial clustering near known star-forming regions, suggesting they are bona fide YSOs that were excluded by initial conservative cuts.
- Generalization: When applied to a new, unlabeled dataset (IPL-4), the model identified an additional 14,426 likely young stars. Of these, 87% were targeted by ABYSS, demonstrating the model's ability to generalize without explicit targeting information.
Identification of Near-IR Youth Signatures: By comparing stacked spectra of young and field stars with matched atmospheric parameters, the authors identified two primary signatures of youth utilized by the classifier:
- Rotational Broadening: Young star spectra exhibit significantly stronger line broadening compared to field stars of the same type, consistent with higher projected rotational velocities ( $v \sin i$ ) typical of young, high-angular-momentum stars.
- Spotted Photosphere Residuals: After accounting for rotational broadening, significant residuals remain in the spectra of young stars. These residuals are attributed to the presence of cool starspots covering a significant fraction of the photosphere (often $>50\%$ ). The authors identified 57 specific spectral features (Table 2) that are likely sensitive to this magnetic activity and the resulting composite temperature profile, rather than simple abundance anomalies.
Catalog Construction: The study produces a catalog of sources identified as likely YSOs, including those with $T_{\text{eff}}$ and $\log g$ comparable to evolved field stars, thereby providing a tool to reduce contamination in photometrically selected YSO candidates.

Significance
The paper claims that this work provides a practical tool for the robust separation of pre-main-sequence stars from more evolved sources in APOGEE data, even when traditional parameters like $\log g$ fail to provide discrimination. By leveraging machine learning to combine weak, distributed spectral signatures—specifically rotational broadening and spot-induced line ratio variations—the study establishes a set of near-IR diagnostics for stellar youth. This capability is critical for analyzing the large volumes of data from SDSS-V, enabling the creation of homogeneous catalogs of spectroscopically selected young sources. Such catalogs are essential for advancing studies of star formation, early stellar evolution, disk evolution, and the kinematic structure of young Galactic populations. The authors acknowledge that while the labels and predictions carry some uncertainty due to the lack of a perfect "ground truth," the systematic distinctness of the stacked spectra validates the physical basis of the classification.

Technical Summary: ABYSS. IV. Identifying signatures of stellar youth in APOGEE spectra

More like this