A machine-learning photometric classifier for massive stars in nearby galaxies II. The catalog

This paper presents a comprehensive catalog of over 1.1 million massive star candidates across 26 nearby galaxies, generated using a machine-learning classifier on Spitzer and Pan-STARRS1 photometry, which identifies nearly 120,500 red supergiants and other evolved massive stars to facilitate studies on mass loss, metallicity effects, and luminosity limits.

The Gist

Imagine the universe as a giant, bustling city. In this city, massive stars are the skyscrapers: they are huge, bright, and they shape the entire neighborhood. But these skyscrapers have a secret life. As they age, they don't just sit there; they go through dramatic phases where they shed massive amounts of their outer layers (like a snake shedding skin, but on a cosmic scale). This "mass loss" is a mystery, and astronomers want to understand it to know how galaxies evolve.

The problem? There are billions of stars, and looking at them one by one with a telescope is like trying to count every grain of sand on a beach by picking them up individually. It takes too long, and for distant galaxies, the stars look too small to see clearly.

This paper is the second part of a project called ASSESS. Think of it as a massive "Star ID" project. Here is the simple breakdown of what they did:

1. The Machine Learning Detective

Instead of hiring a team of astronomers to stare at photos for years, the team built a digital detective (a machine-learning model).

• The Training: They taught this detective by showing it pictures of stars in two nearby galaxies (M31 and M33) where they already knew the answers. They taught it to recognize specific "outfits" (colors) that different types of stars wear.
• The Outfits:
  • Red Supergiants (RSGs): The grumpy, old giants wearing red coats.
  • Yellow Hypergiants: The flashy, unstable stars in yellow suits.
  • Wolf-Rayet stars: The stripped-down, hot stars that have lost their outer layers.
• The Mission: Once trained, they sent this detective to scan 26 different galaxies within 5 million light-years of us.

2. The Great Cleanup (Removing the Foreground)

Before the detective could start, they had to clean the window. When we look at a distant galaxy, we see stars from our own Milky Way floating in front of it, like dust on a camera lens.

• They used data from the Gaia satellite (a cosmic GPS) to figure out which stars were moving with the distant galaxy and which were just "dust" (foreground stars) blocking the view. They filtered out the noise so the detective could see the real targets.

3. The Results: A Massive Catalog

The detective did its job on over 1.1 million sources (stars and other objects).

• The Filter: Not every guess was perfect. The team applied a "confidence filter." They kept the 276,000 guesses that were very sure (about 24% of the total).
• The Big Find: Among these sure bets, they found 120,000 Red Supergiants. That is a huge crowd of aging giants!
• The Metal Test: They tested the detective in galaxies with very low "metal" content (stars with fewer heavy elements, like the early universe). Surprisingly, the detective worked almost as well there as it did in metal-rich places. It's like a detective who can solve crimes in a fancy city and a dusty village with equal skill.

4. The Mystery Stars

The paper highlights a few specific mysteries:

• The "Too Bright" Giants: They found some Red Supergiants that are so bright they shouldn't exist according to current physics rules (the "Humphreys-Davidson limit"). It's like finding a skyscraper that is taller than the laws of architecture say is possible. They might be outliers, or they might mean our rules need updating.
• The Dusty Yellow Stars: They found 159 "Dusty Yellow Hypergiants." Imagine a star that is yellow but is covered in a thick cloud of dust it created itself. These are rare and important because they might be the missing link explaining why we don't see as many massive stars exploding as supernovae as we expect. They might be transforming into something else before they explode.

5. Why This Matters

This paper isn't just a list of numbers; it's a map.

• For the James Webb Space Telescope (JWST): The JWST is the most powerful telescope ever built. It has limited time to look at things. This catalog tells the JWST exactly where to point its eye to find the most interesting, dusty, massive stars to study in detail.
• For the Future: It gives astronomers the largest collection of "spectroscopically confirmed" massive stars (stars we know for sure what they are) outside of our own galaxy and its immediate neighbors.

In a nutshell:
The authors built a smart computer program to sort through millions of stars in nearby galaxies, filtering out the fake ones and the background noise. They created a massive "phone book" of massive stars, found some that break the rules of physics, and identified rare, dusty stars that could solve a decades-old mystery about how massive stars die. This book is now open for other astronomers to use, especially to guide the James Webb Space Telescope.

Technical Summary

Here is a detailed technical summary of the paper "A machine-learning photometric classifier for massive stars in nearby galaxies II. The catalog".

1. Problem Statement

Massive stars play a critical role in galactic evolution through feedback mechanisms (stellar winds, supernovae), particularly in low-metallicity environments relevant to the early Universe. However, understanding episodic mass loss in evolved massive stars is hindered by a lack of large, spectroscopically confirmed samples across diverse galactic environments.

• Challenge: Obtaining spectral classifications for massive stars in distant galaxies is observationally expensive and time-consuming.
• Gap: There is a need for a robust method to classify massive stars in large numbers across a wide range of metallicities (0.07–1.36 $Z_\odot$) and distances (within 5 Mpc) to study evolutionary pathways and the "RSG problem" (the scarcity of luminous Red Supergiants exploding as supernovae).

2. Methodology

The authors applied a machine-learning (ML) classifier developed in their previous work (Paper I) to a sample of 26 nearby galaxies.

• Data Sources:
  • Infrared: Spitzer Space Telescope photometry (3.6, 4.5, 5.8, 8.0, and 24 $\mu$m).
  • Optical: Pan-STARRS1 (PS1) data (g, r, i, z, y bands). For southern galaxies lacking PS1 data, VISTA Hemisphere Survey (VHS) and UKIRT Hemisphere Survey (UHS) were used.
  • Astrometry: Gaia Data Release 3 (DR3) for foreground removal.
• Sample: 26 galaxies within 5 Mpc, covering a metallicity range of 0.07 to 1.36 $Z_\odot$.
• Foreground Cleaning:
  • Utilized Gaia DR3 parallax and proper motion data.
  • Defined galaxy boundaries via visual inspection of optical images (DSS) rather than standard Petrosian radii.
  • Modeled foreground contamination using a regularized cubic spline fitted to sources outside the galaxy boundary, scaled by density ratios.
  • Galactic sources were modeled with a Gaussian distribution; sources exceeding $3\sigma$ of the Gaussian fit were flagged as foreground.
• Classification Model:
  • Algorithm: An ensemble of Support Vector Machines (SVM), Random Forest, and Multi-Layer Perceptron (MLP).
  • Input Features: Color indices derived from optical and IR bands (e.g., $r-i$, $i-z$, $z-y$, $y-[3.6]$, $[3.6]-[4.5]$). Missing data was imputed if up to two features were missing.
  • Classes: Blue Supergiants (BSG), Yellow Supergiants (YSG), Red Supergiants (RSG), B[e] Supergiants (BeBR), Luminous Blue Variables (LBV), Wolf-Rayet (WR), and extragalactic contaminants (GAL).
• Quality Control:
  • Applied a probability cut of $P > 0.66$ (based on the mean of incorrect classifications + $3\sigma$).
  • Applied a band completeness cut of $> 0.6$ (allowing up to 3 missing bands).
  • Sources without Gaia data were retained but not used for foreground cleaning.

3. Key Contributions

• The Catalog: The release of the largest catalog of massive stars and candidates beyond the Local Group, containing 1,147,650 classified sources.
• Robust Subset: A high-confidence subset of 276,657 sources (~24% of the total) meeting strict probability and completeness criteria.
• Spectroscopic Reference: Compilation of the largest spectroscopically confirmed catalog of extragalactic massive stars to date (beyond the Milky Way and Magellanic Clouds), comprising 5,273 sources from 83 literature works.
• Methodological Advancement: Demonstration of an ML classifier's ability to generalize to low-metallicity environments ($\sim0.1 Z_\odot$) without explicit retraining, and the implementation of a rigorous Bayesian approach to evaluate classifier success rates against literature data.

4. Key Results

• Classification Statistics (Robust Subset):
  • RSGs: 120,479 (~44% of robust sample).
  • WR Stars: 150,151 (Note: The paper acknowledges a high false-positive rate for WR predictions, requiring careful follow-up).
  • YSGs: 2,082.
  • BSGs: 616.
  • BeBRs: 72.
  • LBVs: 60.
  • GAL (Contaminants): 3,197.
• Classifier Performance:
  • Metallicity: The classifier performs robustly even at very low metallicities ($\sim0.07 Z_\odot$), with success rates remaining relatively flat across the metallicity range.
  • Distance: Accuracy is high for distances $<1.5$ Mpc. A slight decrease in accuracy is observed beyond $\sim3$ Mpc, attributed to Spitzer's spatial resolution limits and source confusion.
  • Success Rate: For well-populated galaxies (M31, M33), success rates are high ($\sim80-90\%$). For galaxies with small literature samples, rates vary but are corrected via Bayesian priors to realistic ranges (60–94%).
• Luminous RSGs:
  • Identified 21 luminous RSGs with $\log(L/L_\odot) \ge 5.5$.
  • Found 6 extreme RSGs in M31 with $\log(L/L_\odot) \ge 6$, challenging the Humphreys-Davidson (HD) limit. Some may be foreground stars or possess dense circumstellar envelopes contributing to luminosity estimates.
• Dusty Yellow Supergiants (YSGs):
  • Identified 159 dusty YSGs (based on IR color $[3.6]-[4.5] > 0.24$ mag).
  • These are potential candidates for Yellow Hypergiants (YHGs), which may represent the evolutionary link between RSGs and supernovae, addressing the "RSG problem."
• Metallicity Trends:
  • Observed a decrease in WR fraction at lower metallicities (consistent with wind-driven mass loss theory).
  • Observed an increase in BSG and YSG fractions at lower metallicities, likely due to extended main-sequence phases and faster rotation in low-Z environments.

5. Significance

• JWST Target Selection: The catalog serves as a primary resource for selecting targets for the James Webb Space Telescope (JWST), particularly for studying massive stars in low-metallicity environments.
• Evolutionary Pathways: The identification of luminous RSGs and dusty YSGs provides critical data to refine models of massive star evolution, mass loss, and the "RSG problem."
• Future Research: The catalog enables statistical studies of massive star populations across a wide metallicity range, which was previously impossible due to the lack of large, classified samples.
• Data Availability: The full catalog and the spectroscopic reference list are made available via the CDS (Centre de Données astronomiques de Strasbourg), facilitating further independent analysis and follow-up spectroscopy.

In conclusion, this paper successfully bridges the gap between photometric surveys and spectroscopic classification, providing a massive, statistically significant dataset that advances our understanding of massive star evolution in the local Universe and beyond.