🔭 astrophysics

Mapping the Inner Milky Way with Infrared-Derived Distances to AGB Stars

This paper presents a supervised machine-learning model using XGBoost to estimate statistical distances for over 36,000 oxygen-rich AGB stars from AKARI infrared data, successfully mapping the inner Milky Way's structure and revealing that long-period Mira variables are superior tracers of the Galactic bar compared to their shorter-period counterparts.

Original authors: Rajorshi Bhattacharya, Subhajit Kar, Loránt O. Sjouwerman, Anupam Bhardwaj

Published 2026-08-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rajorshi Bhattacharya, Subhajit Kar, Loránt O. Sjouwerman, Anupam Bhardwaj

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

Imagine the Milky Way not as a flat, quiet disk, but as a bustling, dusty city where the streets are so crowded with gas and grime that you can't see the buildings on the other side. Astronomers have long wanted to map this inner city to understand how our galaxy was built and how it grows, but they face a massive problem: the "fog" of interstellar dust blocks visible light, like trying to navigate a thick fog bank with a flashlight that only works in the daytime. To see through this fog, scientists use infrared light, which is like a special pair of night-vision goggles that can pierce the dust. However, even with these goggles, there's a tricky part: knowing exactly how far away a star is. Usually, astronomers measure distance by watching a star's tiny shift in position as Earth orbits the Sun (like holding your thumb up and closing one eye, then the other), but for these dusty, bloated old stars, that method is too fuzzy to be useful. Without accurate distances, the galaxy looks like a flat, confusing smear rather than a 3D structure with bars, bulges, and spirals.

This is where a team of astronomers stepped in with a digital detective tool: machine learning. They decided to teach a computer to guess the distance of these dusty stars by looking at their colors, much like how a real estate agent might estimate a house's value based on its paint color and neighborhood, even without walking inside. They focused on a specific type of "retired" star called an Asymptotic Giant Branch (AGB) star. These are massive, aging stars that have puffed up and are shedding their outer layers, creating a thick, dusty cocoon around them. Because they are so bright in infrared light, they are perfect for mapping the dusty inner galaxy, but figuring out their distance has been a headache. The researchers used a massive dataset of these stars to train a smart algorithm, teaching it the relationship between a star's infrared brightness and its actual distance. Once the computer learned the pattern, they let it loose on a huge new list of stars from the AKARI satellite survey, effectively turning a flat, 2D map of the galaxy into a detailed 3D model.

The paper's main finding is that this machine-learning approach works remarkably well. By feeding the computer data from two different infrared surveys (AKARI and 2MASS), the model successfully estimated distances for over 36,000 of these dusty stars, covering a range from 0.5 to 20 kiloparsecs (about 1,600 to 65,000 light-years). The model is statistically robust, with an average error of about 6% on its test runs, and when applied to the full sample, the total error margin stays within a manageable 36%. This allows the team to finally "see" the structure of the inner Milky Way in three dimensions. They discovered that the galaxy's central bulge isn't just a round ball of stars; it has a distinct, elongated "bar" shape, like a peanut or a football.

Crucially, the paper argues against the idea that this structure is just a random blur or an illusion caused by the dust. Instead, the data suggests that the galaxy's shape is real and can be traced by specific types of stars. The authors explicitly rule out the notion that their distance estimates are just random guesses; they show that the distances align with other known methods, such as the relationship between a star's pulsation period and its brightness (a bit like how the pitch of a drum might tell you its size). They also argue that the "fog" of dust isn't tricking them into thinking stars are farther away than they are; the patterns they see are consistent with the physical reality of the galaxy.

One of the most playful and insightful discoveries in the paper is that not all these old stars tell the same story. The team found that the "younger" and more massive of these aging stars (those with longer pulsation periods, greater than 400 days) are the ones that clearly trace the galaxy's central bar. They cluster tightly along the elongated shape, acting like neon signs outlining the galaxy's spine. In contrast, the "older" stars (with shorter periods, less than 400 days) seem to have drifted apart over time, forming a smoother, rounder cloud that doesn't show the bar as clearly. This suggests that the bar is a relatively young feature in the galaxy's history, and only the younger, more massive stars are still tightly bound to its structure.

The paper also maps the "height" of the galaxy. They found that the central bulge is vertically thicker and more "puffed up" than the surrounding disk, with stars in the bulge spreading out about 30% more in the vertical direction. This difference in shape helps confirm that the bulge and the disk are distinct neighborhoods with different histories. While the authors are careful to note that these are statistical estimates and not perfect measurements for every single star, the collective picture is clear: the inner Milky Way has a barred structure, and long-period Mira variables are the best guides to see it. The study doesn't claim to have solved every mystery of the galaxy, nor does it suggest that the distances are perfect for individual stars, but it successfully demonstrates that machine learning can cut through the cosmic dust to reveal the galaxy's hidden 3D architecture.

=== DRAFT ===
Technical Summary: Mapping the Inner Milky Way with Infrared-Derived Distances to AGB Stars

Problem Statement
Understanding the formation and evolution of the inner Milky Way (MW) requires accurate distance estimates to evolved stellar populations, specifically Asymptotic Giant Branch (AGB) stars. However, direct astrometric distances (e.g., from Gaia) are largely unavailable or highly uncertain for these sources due to severe interstellar extinction, large angular diameters, variability-induced astrometric jitter, and complex circumstellar environments. Consequently, the majority of AGB stars in the inner Galaxy, particularly those in the dust-obscured Galactic plane and bulge, lack reliable distance measurements. Existing complementary methods, such as kinematic distances and Period-Luminosity (P-L) relations, suffer from significant uncertainties in regions of non-circular motion or for dustiest subtypes where mass loss and metallicity distort observed magnitudes.

Methodology
To address these limitations, the authors developed a supervised machine-learning (ML) regression model to estimate statistical distances for oxygen-rich (O-rich) AGB stars using multi-band infrared (IR) photometry.

Data Selection: The study utilizes the AKARI mid-infrared survey (9 µm and 18 µm bands), which offers deeper sensitivity and higher angular resolution (5″–6″) compared to the previous MSX survey used by the Bulge Asymmetries and Dynamical Evolution (BAaDE) project. A clean sample of O-rich AGB candidates was selected using a conservative color criterion ( $-0.6 < [9]-[18] < 0.6$ ) and cross-matched with 2MASS near-IR photometry. This resulted in a final sample of 36,134 sources (the "AKARI AGB sample") distinct from the training set.
Training Set: The model was trained on a subset of 10,112 MSX-selected AGB stars from a previous study (R. Bhattacharya et al. 2024) that possessed independently derived, physically motivated Spectral Energy Distribution (SED)-based distances.
Model Architecture: An XGBoost (Extreme Gradient Boosting) regression model was employed to map apparent IR magnitudes (2MASS $J, H, K_s$ and AKARI 9 µm, 18 µm) directly to distance. The model utilized apparent magnitudes without explicit extinction corrections to avoid introducing systematic uncertainties from ancillary data.
Validation: The model was validated using an 80/20 train-test split and 5-fold cross-validation. Performance was assessed using $R^2$ , Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE).

Key Results

Model Performance: The XGBoost model achieved an $R^2 \approx 0.98$ on both training and independent test sets, with a Mean Absolute Percentage Error (MAPE) of approximately 6% on the test set. The model demonstrated robustness across the full distance range (0.5–20 kpc) with no significant systematic bias or overfitting.
Distance Catalog: The trained model was applied to the full AKARI AGB sample, yielding statistical distance estimates for over 36,000 sources. The total error margin for these estimates is approximately 35.5%, combining the ML methodological uncertainty (~~6%) with the inherent uncertainty of the SED-based training distances (~~35%) in quadrature.
Validation Against Independent Data:
- Gaia Parallaxes: Comparison with Gaia DR3 parallaxes for sources at higher Galactic latitudes ( $|b| > 5^\circ$ ) showed a strong inverse correlation ( $\rho = -0.72$ ). The ML distances were generally consistent with Gaia, though Gaia tended to underestimate distances for fainter sources ( $G > 12$ ) beyond ~2 kpc, a known limitation.
- P-L Relations: Comparisons with various P-L relations (OGLE, MSX-based, and Gaia-calibrated) showed that the ML distances align well with Galactic-calibrated relations, particularly mid-IR P-L relations for Galactic Miras. Discrepancies with Large Magellanic Cloud (LMC)-calibrated relations were attributed to metallicity differences and distance scale offsets.
Galactic Structure Analysis:
- Barred Morphology: The spatial distribution of the derived distances reveals a clear barred structure in the Galactic bulge. Long-period Mira variables ( $P > 400$ days) preferentially trace this elongated, triaxial bar, whereas short-period Miras ( $P \leq 400$ days) exhibit a smoother, more spheroidal distribution consistent with an older, dynamically mixed population.
- Vertical Structure: Analysis of the dimensionless vertical dispersion ( $\tilde{\sigma}_z$ ) indicates that the Galactic bulge has a larger vertical dispersion (approximately 30% larger) and is dynamically hotter than the surrounding disk. The transition between the thick bulge and the thinner disk is clearly resolved, validating the reliability of the distance scale for population-level studies.

Significance and Claims
The paper claims that IR photometry-derived statistical distances can effectively recover large-scale Galactic structures in heavily obscured regions where direct astrometry fails. The primary contribution is the expansion of the distance catalog for AGB stars in the inner MW from a few thousand to over 36,000 sources, significantly extending the accessible volume into the Galactic plane and bulge.

The authors assert that their ML-based approach provides a homogeneous and statistically robust method for mapping the inner Galaxy without requiring time-consuming flux monitoring or explicit extinction corrections. The results establish long-period Mira variables as efficient tracers of younger, massive stellar populations associated with the Galactic bar, while confirming that short-period Miras trace a more spheroidal, older component. The study concludes that these statistical distances are suitable for population-level analyses of Galactic structure and kinematics, offering a scalable solution for future surveys (e.g., JWST, Roman, Rubin/LSST) to constrain the chemodynamical evolution of the inner Milky Way.

More like this