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

The paper Mapping the Inner Milky Way with Infrared-Derived Distances to AGB Stars develops a supervised machine-learning model to estimate statistical distances to oxygen-rich asymptotic giant branch (AGB) stars in the inner Milky Way, using multi-band infrared photometry. The authors train an XGBoost regression model on a sample of AGB stars with previously derived SED-based distances, mapping 2MASS (J, H, Ks) and AKARI (9 µm, 18 µm) apparent magnitudes to distance. The model achieves a mean absolute percentage error (MAPE) of approximately 6% on an independent test set, with R2 ≈ 0.98 for both training and test sets, indicating strong performance without significant overfitting. The model is applied to a sample of 36,134 AKARI-selected O-rich AGB candidates, yielding distance estimates between 0.5 and 20 kpc with a total error margin of approximately 36%, which combines the 35% uncertainty of the SED-based training distances with the 6% model uncertainty.

The AKARI sample is constructed using a color criterion of −0.6 < [9] − [18] < 0.6 in AKARI bands, cross-matched with 2MASS within a 1″ radius, and restricted to Galactic latitudes b < 5°. This selection yields nearly twice as many sources as the previous MSX/BAaDE selection, with improved coverage of both the near and far sides of the Galactic bulge. The sample purity is estimated at 97.6%, with contamination dominated by YSOs (2.4%) and a small fraction of C-rich AGB stars (0.29%).

The paper validates the ML-derived distances through multiple independent comparisons. First, comparison with Gaia parallaxes for 166 sources at higher Galactic latitudes shows a strong inverse correlation (Spearman ρ = −0.72), with the brightest sources (G < 8) showing the tightest agreement, while fainter sources exhibit larger scatter consistent with known Gaia biases for distant AGB stars. Second, comparison with period-luminosity (P-L) relations reveals that distances derived from Galactic-calibrated mid-IR P-L relations (M. O. Lewis et al. 2023b) agree within 12% of the ML distances, while LMC-calibrated near-IR P-L relations (P. Iwanek et al. 2023) systematically underestimate distances by 36%, consistent with known metallicity and distance-scale differences. Third, bolometric magnitudes derived from ML distances for 143 cross-matched OH/IR sources peak at slightly brighter values (0.4 mag) compared to a Gaia-based catalog, which is attributed to the different stellar populations probed.

The spatial distribution of the AGB sample reveals a bar-like structure in the inner Galaxy, with a clear near-far asymmetry in distance distributions toward positive versus negative longitudes, consistent with an inclined triaxial bar. The paper also investigates the vertical structure of the inner Galaxy using a dimensionless vertical dispersion parameter, finding that the bulge exhibits higher and nearly uniform vertical dispersion (σ̃z ≈ 0.40) compared to the disk (σ̃z ≈ 0.30), with a smooth transition between the two components. This is consistent with a dynamically hotter, vertically thicker bulge and a thinner, more uniform disk.

A key result concerns the period-dependent spatial distribution of Mira variables. Long-period Miras (P > 400 days) preferentially trace an elongated barred morphology, while short-period Miras (P ≤ 400 days) show a smoother, more spheroidal distribution. This is consistent with younger, more massive AGB populations being associated with the bar, while older populations are more dynamically mixed. The paper notes that pulsation period is not an input feature in the ML model, so this period-dependent difference emerges only after independently estimated distances are separated by period, supporting the physical interpretation. The short-period Miras show a broad, potentially bimodal distance distribution with a far-side enhancement, which the authors cautiously interpret as possibly affected by projection effects, sample selection, or systematic uncertainties, and defer detailed investigation to future work.

The paper concludes that ML-based distance estimation from IR photometry can effectively scale up distance information for large samples of dust-obscured AGB stars, recovering large-scale Galactic structures such as the bar, and establishing long-period Mira variables as efficient tracers of younger stellar populations in the bulge and disk. The authors suggest that combining these statistical distances with BAaDE maser kinematics, enhanced extinction modeling, and upcoming IR time-domain surveys will improve chemodynamical constraints on the formation and evolution of the inner Milky Way.

Improvements to AI Systems:

1. Domain-Adapted Regression for High-Dimensional Photometric Data: Enhance XGBoost or similar gradient-boosting frameworks with custom loss functions that explicitly model heteroscedastic uncertainty (e.g., quantile regression or Gaussian negative log-likelihood) to output both distance and per-source confidence intervals, rather than point estimates. This would allow the AI to flag low-reliability predictions for sources near the Galactic plane or with high extinction.

2. Causal-Aware Feature Engineering for Stellar Population Classification: Integrate a pre-trained classifier that separates O-rich AGB, C-rich AGB, and YSOs using the same photometric bands, then use the class probabilities as auxiliary inputs to the distance regressor. This reduces contamination bias and improves distance accuracy for mixed populations, as demonstrated by the 2.4% YSO contamination in the sample.

3. Cross-Validation with Multi-Scale Priors: Implement a two-stage AI pipeline: (a) a photometric-distance model as in the paper, and (b) a Bayesian refinement layer that incorporates independent distance anchors (e.g., Gaia parallaxes, period-luminosity relations) to correct systematic offsets—like the 36% underestimation from LMC-calibrated relations—by learning a metallicity- or extinction-dependent correction term.

4. Transfer Learning for Time-Domain Surveys: Train the model on static 2MASS/AKARI photometry, then fine-tune on simulated or real time-series data (e.g., from upcoming IR surveys like WISE/NEOWISE or LSST) to predict distances for variable AGB stars using pulsation-phase-averaged magnitudes, leveraging the period-dependent spatial structure (long vs. short period Miras) as a latent feature.

5. Uncertainty-Aware Spatial Mapping for Galactic Structure: Use the model’s predicted distances and their error margins to generate probabilistic 3D density maps of the inner Galaxy (e.g., via Gaussian process or normalizing flows), enabling automatic detection of bar-like overdensities, vertical dispersion gradients (σ̃z ≈ 0.40 vs 0.30), and near-far asymmetries without manual binning.

6. Self-Supervised Pretraining on Unlabeled IR Catalogs: Pretrain a transformer or CNN on a large unlabeled set of 2MASS/AKARI color-magnitude diagrams to learn robust representations of dust-obscured stellar populations, then fine-tune on the labeled AGB sample. This improves generalization to fainter sources or higher extinction regions where SED-based training distances are sparse.

What the Improved AI System Can Do:

• Provide reliable, uncertainty-calibrated distances for millions of dust-obscured AGB stars in the Milky Way, with per-source error bars that reflect both model and data noise, enabling robust statistical analyses of Galactic structure.

• Automatically separate stellar populations (O-rich vs. C-rich AGB, YSOs) and distance estimates simultaneously, reducing contamination and improving purity in large surveys.

• Self-correct systematic biases by integrating heterogeneous distance anchors (Gaia, P-L relations) in real time, yielding distances that are consistent across different metallicity regimes and distance scales.

• Generate time-resolved 3D maps of the inner Galaxy from upcoming IR time-domain surveys, tracing the bar’s morphology and the vertical structure of the bulge/disk, and distinguishing young (long-period) vs. old (short-period) AGB populations dynamically.

• Flag anomalous sources (e.g., possible YSOs or C-rich stars) with high confidence, aiding follow-up spectroscopic or maser observations (e.g., BAaDE) for chemodynamical studies.

• Scale to other Galactic regions or external galaxies by fine-tuning on local training sets, enabling distance estimation for obscured stellar populations beyond the inner Milky Way.

The structure and evolution of the Milky Way (MW) can be traced with distance estimates to the evolved stellar populations in the inner galactic region. However, direct astrometric distances remain unavailable or highly uncertain for the majority of these sources due to instrumental limitations, large angular diameters, and complex variability. In this work, we develop a supervised machine-learning model to estimate statistical distances to oxygen-rich asymptotic giant branch (AGB) stars selected from the AKARI mid-infrared survey. We build an XGBoost regression model that maps multi-band IR photometry to distance using a training set of AGB stars with previously derived SED-based distances achieving a mean absolute percentage error (MAPE) of 6% on an independent test set. Distance estimates for over 36,000 AGB sources are obtained within a 36% total error margin, greatly expanding distance coverage (0.5-20 kpc) for dust-obscured AGB populations in the Galactic plane. We find good agreement with reported distances for Galactic Mira variables and independent period-luminosity relations. Utilizing the expanded distance set, we further investigate the spatial distribution of Mira variables in the Galactic bulge and disk. Longer-period Miras preferentially trace the bulge's barred morphology compared to their shorter-period counterparts that populate the disk. We also find roughly constant, but differing, relative scale heights for the bulge and disk, with the bulge vertical dispersion about 30% larger. These findings show that IR photometry-derived statistical distances can recover large-scale Galactic structures and establish long-period Mira variables as efficient tracers of stellar populations in heavily obscured regions of the MW.