Astro+ database. I. Description and first results

The paper presents the Astro+ database, a new multiwavelength repository for massive star spectra, and its fully automated analysis tool, HiLineThere. The database currently contains over 4,000 stellar spectra, mostly OB and red supergiant (RSG) spectra from public catalogs (OWN, IACOB, GOSSS) and the authors' own observations. The tool is designed to handle the large influx of spectroscopic data from current and upcoming surveys (e.g., WEAVE) that exceed human processing capabilities.

The analysis pipeline uses a hierarchical classification logic with three paths:

• Blue path (hot OB stars, T eff > 15,000 K): Uses a grid of over 46,000 FASTWIND models (solar metallicity, NLTE, spherical geometry, mass loss, line blanketing) to derive T eff, log g, helium abundance, wind strength, and wind velocity exponent via reduced chi-squared minimization. The grid covers T eff from 10,000 to 55,000 K and log g from 2.6 to 4.3 dex.

• Yellow path (intermediate stars, 7,500–15,000 K): Uses KURUCZ/ATLAS9 models via a subroutine called HiBandThereY to bridge the gap between hot and cool regimes, deriving parameters down to 7,500 K.

• Red path (cool stars, including RSGs): Uses MARCS models. A subroutine called HiBandThere determines radial velocity (Vr) by cross-correlating with 64 cool stellar models (T eff 2000–8000 K). Then SteParSyn, a Bayesian code using principal component analysis on two MARCS model networks (1 M sun and 15 M sun), derives T eff, log g, and metallicity.

The tool automatically detects diagnostic lines (Balmer series, He I, He II, Si, C, N, Mg, Fe lines), determines Vr, projected rotational velocity (V sin i) via Fourier transform technique, and macroturbulence (ζ) via Monte Carlo simulations (100 iterations). It also handles line blending (e.g., He I 6678/He II 6682), removes contaminants (cosmic rays, interstellar lines), and excludes Hα from the fit when wind effects are detected.

Validation results:

• O-type stars (sample from Holgado et al. 2018): Differences within expected error ranges: T eff 800 K, log g 0.1 dex, V sin i 10 km/s. A notable outlier (HD 105056, an ON-type supergiant with anomalous CNO abundances) was resolved when the code was allowed to freely select diagnostic lines instead of using a fixed subset.

• Early-B stars (sample from Nieva & Przybilla 2014): Consistent results during the transition to lower temperatures, with good correlation with spectral types.

• Red supergiants: Vr differences within 7 km/s compared to Dorda et al. (2018), and T eff differences 300 K for a small sample from Negueruela et al. (2018).

• Systematic errors: Tests using TLUSTY models as input spectra (at resolutions R = 5000, 25000, 48000) showed T eff differences below 2,000 K for O-type stars, with larger discrepancies in log g for high-gravity models due to resolution effects.

The paper concludes that Astro+ provides a powerful, versatile, and user-friendly platform for homogeneous, fully automated analysis of massive star spectra, free from subjective human biases. Future work (Paper II) will address systematic validation of the yellow path and extension to other metallicity regimes. The database is publicly accessible at https://astroplus.ua.es/.

Improvements to AI Systems:

1. Hierarchical Multi-Model Classification Pipeline

• Implement a tiered decision-tree system where the AI automatically selects among multiple physical model grids (FASTWIND, KURUCZ/ATLAS9, MARCS) based on preliminary temperature estimates.

• Capability: The AI can robustly analyze stellar spectra across the full temperature range (2,000–55,000 K) without manual intervention, adapting its underlying physics to the star's regime.

2. Automated Diagnostic Line Selection with Anomaly Handling

• Use a dynamic feature-selection algorithm that identifies the most informative spectral lines for each spectrum, rather than a fixed set. When outliers (e.g., anomalous CNO abundances) cause poor fits, the AI re-runs with a relaxed line-selection criterion.

• Capability: The AI avoids systematic biases from unusual chemical compositions and improves accuracy for rare or exotic stellar types (e.g., ON-type supergiants).

3. Bayesian Parameter Estimation with Principal Component Analysis (PCA)

• Integrate a Bayesian inference module (like SteParSyn) that reduces high-dimensional model grids via PCA, then samples posterior distributions for T eff, log g, and metallicity.

• Capability: The AI provides uncertainty quantification for each derived parameter, not just point estimates, enabling more reliable astrophysical conclusions.

4. Multi-Component Line Blending and Contaminant Removal

• Train a convolutional neural network to detect and deblend overlapping spectral lines (e.g., He I 6678/He II 6682) and to flag cosmic rays or interstellar absorption features for automatic masking.

• Capability: The AI can process noisy, low-resolution, or contaminated spectra without human cleaning, increasing throughput for large surveys.

5. Wind-Strength and Hα Exclusion Logic

• Implement a rule-based trigger that detects strong stellar winds (via Hα profile asymmetry) and automatically excludes Hα from the fitting process, while still using it to derive wind parameters (terminal velocity, mass-loss rate).

• Capability: The AI accurately characterizes hot, massive stars with strong mass loss, avoiding systematic errors in gravity and temperature from wind-contaminated lines.

6. Resolution-Adaptive Error Correction

• Train a meta-model that predicts systematic biases in log g and T eff as a function of spectral resolution (R = 5,000–48,000) and signal-to-noise ratio, then applies a correction factor to the output.

• Capability: The AI maintains consistent accuracy across heterogeneous datasets from different telescopes/surveys, enabling homogeneous comparisons.

7. Fully Automated Radial Velocity and Rotational Broadening Estimation

• Use cross-correlation with a library of template spectra (64 models for cool stars) and a Fourier transform technique for V sin i, followed by Monte Carlo simulations (100 iterations) for macroturbulence.

• Capability: The AI outputs complete kinematic and line-broadening parameters without human input, suitable for real-time processing of streaming survey data.

8. Public Database Integration and Continuous Learning

• Build an API that connects the AI to the Astro+ database (4,000+ spectra) for online validation, and implement a feedback loop where new high-confidence results are used to fine-tune the model grids or neural networks.

• Capability: The AI improves over time as more spectra are analyzed, adapting to new stellar populations or instrument characteristics.

9. Subjective Bias Elimination via Homogeneous Processing

• Replace human-in-the-loop decision points with a deterministic, fully automated pipeline that logs every step (line selection, model choice, fitting criteria) for reproducibility.

• Capability: The AI produces scientifically consistent results across different users and studies, eliminating inter-observer variability in stellar parameter derivation.

10. Scalable Parallel Processing for Large Surveys

• Design the pipeline to run on distributed computing (e.g., Apache Spark or Kubernetes) with batch processing of thousands of spectra, using the hierarchical logic to allocate computational resources (e.g., expensive FASTWIND grids only for hot stars).

• Capability: The AI can handle upcoming survey data (e.g., WEAVE) in near-real-time, processing millions of spectra within survey timelines.

The vast amounts of spectroscopic data for massive stars provided by previous and existing instruments on ground-based and space-based telescopes have saturated our capability to process them by human inspection routines. Consequently, there is a pressing need for fully automatic machine-assisted tools to help handle incoming data. To this end, we present the development of a massive star spectroscopic interactive database, Astro+. We aim to provide users with a reliable, versatile, and user-friendly platform that will be significant for understanding massive stars and set an important precedent for future open-access astrophysical research. This tool allows authorized users to upload their own spectra and, by using the fully automated tool HiLineThere, a Python-based program, it can homogeneously derive basic stellar parameters, such as v rad, v sin i, Teff and log g for massive OB-type stars, using a solar-metallicity grid of FASTWIND models, and v rad, [Fe/H], Teff and log g for red supergiant stars, in a completely autonomous way. Here we present the first results of the tool HiLineThere on optical spectra for OB-type stars and red supergiants. We compare the output of our analysis for OB-type stars with literature values for a large sample of well-studied objects, finding differences within the expected error ranges: Teff 800 K, log g 0.1 dex, and v sin i 10 km/s. Preliminary tests on early-B stars also show consistent results during the transition to lower temperatures. For red supergiants, we find differences within 7 km/s in v rad and 300 K in Teff for two test samples.