Datasets:

Dedulek
/

gaia-eb-teff-datasets

text stringlengths 0 107
# stars_types_with_best4_predictions.fits - Description

## Overview

This catalog contains effective temperature predictions for 2.1 million eclipsing binary stars.
It merges the base stars_types.dat catalog with ML predictions using a "best-of-four" ensemble
approach that selects the prediction with the lowest uncertainty for each object.

Creation Date: 2025-12-18
Total Objects: 2,145,310
Objects with Teff: 2,085,712 (97.2%)
Temperature Range: 2737 - 38456 K

## Temperature Sources

1. Gaia GSP-Phot: 1,251,127 objects (58.3%)
- High-quality spectrophotometric temperatures from Gaia DR3
- No uncertainty estimates provided

2. ML Predictions: 834,585 objects (38.9%)
- Best-of-four ensemble (selects lowest uncertainty)
- Four models compared:
* teff_only: Gaia photometry only (g, BP, RP, BP-RP)
* teff_logg: Gaia photometry + log(g) with uncertainty propagation
* teff_cluster: Gaia photometry + cluster probabilities
* teff_flag1: Gaia photometry (trained on flag 1 high-quality sources)
- Mean uncertainty: 203 K

3. No Prediction: 59,598 objects (2.8%)
- Objects without Gaia Teff and outside ML training domain

## Quality Flags

Quality assessment based on temperature source and uncertainty:

- A: Gaia GSP-Phot temperature (highest quality) - 1,251,127 objects (58.3%)
- B: ML prediction with uncertainty < 300 K (high confidence) - 736,974 objects (34.4%)
- C: ML prediction with uncertainty < 500 K (medium confidence) - 64,031 objects (3.0%)
- D: ML prediction with uncertainty >= 500 K (low confidence) - 33,580 objects (1.6%)
- X: No temperature available - 59,598 objects (2.8%)

## Model Selection Distribution (ML predictions only)

Best-of-four ensemble selects the model with lowest uncertainty for each object:

- teff_cluster: 160,931 objects (19.3%)
- teff_flag1: 243,801 objects (29.2%)
- teff_logg: 149,829 objects (18.0%)
- teff_only: 280,024 objects (33.6%)

## Uncertainty Statistics (ML predictions only)

- Mean: 203 K
- Median: 168 K
- Std Dev: 144 K
- Min: 17 K
- Max: 5372 K
- 25th percentile: 124 K
- 75th percentile: 239 K

## Column Descriptions

\| Column \| Type \| Unit \| Description \|
\|--------\|------\|------\|-------------\|
\| source_id \| int64 \| - \| Gaia DR3 source identifier \|
\| ra \| float64 \| deg \| Right Ascension (J2000) \|
\| dec \| float64 \| deg \| Declination (J2000) \|
\| period \| float64 \| d \| Orbital period \|
\| teff_gaia \| float64 \| K \| Effective temperature from Gaia GSP-Phot (null if unavailable) \|
\| binary_type \| str \| - \| Binary type (D=detached, C=overcontact) \|
\| amplitude \| float64 \| mag \| Light curve amplitude \|
\| teff_predicted \| float64 \| K \| ML predicted temperature from best-of-four (null if no prediction) \|
\| teff_uncertainty \| float64 \| K \| ML prediction uncertainty (null for Gaia temperatures) \|
\| teff_final \| float64 \| K \| Final temperature (Gaia if available, else ML, null if neither) \|
\| teff_source \| str \| - \| Temperature source (Gaia, teff_only, teff_logg, teff_cluster, teff_flag1, none) \|
\| quality_flag \| str \| - \| Quality flag (A/B/C/D/X, see above) \|

## Usage Examples

### Python (Astropy)

```python
from astropy.table import Table

# Load catalog
catalog = Table.read('stars_types_with_best4_predictions.fits')

# Filter by quality
high_quality = catalog[catalog['quality_flag'] <= 'B'] # Gaia or low-uncertainty ML
print(f"High-quality objects: {len(high_quality):,}")

# Access temperatures
teff = catalog['teff_final']
uncertainty = catalog['teff_uncertainty']

# Filter by temperature range
cool_stars = catalog[(catalog['teff_final'] > 3000) & (catalog['teff_final'] < 5000)]
```

### Python (Polars)

End of preview. Expand in Data Studio

Gaia Eclipsing Binary Effective Temperature Dataset

This dataset provides effective temperature (Teff) predictions for 2.1 million eclipsing binary stars using a best-of-four ensemble approach combining Random Forest models trained on Gaia DR3 photometry.

Dataset Summary

Total Objects: 2,145,310 eclipsing binaries Teff Coverage: 97.2% (2,085,712 objects with temperature estimates) High Quality (A+B flags): 92.7% (1,933,634 objects) Data Sources: Gaia DR3, Pan-STARRS, 2MASS

Best-of-Four Ensemble

The dataset uses a best-of-four ensemble that selects the prediction with the lowest Random Forest uncertainty for each object from four models:

Teff-only model (colors only, corrected Gaia temperatures)
Teff+logg model (colors + surface gravity)
Teff+clustering model (colors + cluster features)
Flag 1 model (trained on highest quality Gaia sources)

Performance:

Mean uncertainty: 203 K (22.8% improvement over single best model)
Model selection: 29.2% flag 1, 25.6% colors-only, 23.8% +logg, 21.4% +clustering

Files

Main Catalog

catalogs/stars_types_with_best4_predictions.fits (206 MB)
- Complete catalog with best-of-four Teff predictions
- Columns: source_id, ra, dec, teff_final, teff_uncertainty, quality_flag, best_model
- See accompanying _DESCRIPTION.txt for full schema

Photometry

photometry/eb_unified_photometry.parquet (69 MB)
- Gaia DR3 + Pan-STARRS + 2MASS photometry for all eclipsing binaries
- Includes magnitudes, colors, and photometric uncertainties

Correction Coefficients

correction/teff_correction_coeffs_deg2.pkl
- Polynomial coefficients for correcting Gaia GSP-Phot temperatures (>10,000K)
- Used during model training to correct systematic underestimation

Quality Flags

Flag	Description	Uncertainty	Count	Percentage
A	Gaia GSP-Phot original	N/A	1,249,946	58.3%
B	ML prediction (< 300K)	< 300K	683,688	31.9%
C	ML prediction (300-500K)	300-500K	152,078	7.1%
D	ML prediction (≥ 500K)	≥ 500K	6,000	0.3%
X	No temperature available	N/A	53,598	2.5%

High Quality (A+B): 92.7% of catalog ML Predictions (B+C+D): 39.2% (841,766 objects)

Usage Example

from astropy.table import Table
import pandas as pd

# Load FITS catalog
catalog = Table.read('stars_types_with_best4_predictions.fits')

# Convert to pandas for analysis
df = catalog.to_pandas()

# Filter high-quality predictions
high_quality = df[df['quality_flag'].isin(['A', 'B'])]

# Get objects with low uncertainty ML predictions
low_unc_ml = df[(df['quality_flag'] == 'B') & (df['teff_uncertainty'] < 200)]

print(f"High quality sources: {len(high_quality):,}")
print(f"Low uncertainty ML predictions: {len(low_unc_ml):,}")

Model Information

Pre-trained models are available in the companion repository: Dedulek/gaia-eb-teff-models

Flag 1 Model (Primary contributor to ensemble)

Model: rf_gaia_colors_flag1_20251218_094937.pkl
Training: 711,666 high-quality Gaia sources (flag 1)
Features: 6 Gaia colors + bands (g, bp, rp, bp_rp, g_rp, g_bp)
Performance: MAE 195K, R² 0.9126
Contribution: Selected for 29.2% of ML predictions

Citation

If you use this dataset, please cite:

@dataset{gaia_eb_teff_2024,
  title={Gaia Eclipsing Binary Effective Temperature Dataset},
  author={Your Name/Team},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/Dedulek/gaia-eb-teff-datasets}}
}

Data Processing

This dataset was created using a configurable ML pipeline:

Training Data: Gaia DR3 GSP-Phot temperatures with polynomial correction (>10,000K)
Features: Gaia photometry (g, bp, rp bands) + derived colors
Model: Random Forest with full tree uncertainty estimation
Ensemble: Best-of-four selection based on RF uncertainty
Quality Control: Systematic validation against spectroscopic surveys

License

This dataset is released under CC BY 4.0. You are free to share and adapt with attribution.

Related Resources

Models: Dedulek/gaia-eb-teff-models
Gaia Archive: https://gea.esac.esa.int/archive/
Pan-STARRS: https://panstarrs.stsci.edu/

Updates

2024-12-18: Updated to best-of-four ensemble (mean uncertainty: 203K)
Previous: best-of-three ensemble (mean uncertainty: 263K)

Contact

For questions or issues, please open an issue in the dataset repository or contact the authors.

Downloads last month: 27

Size of downloaded dataset files:

8.95 kB

Size of the auto-converted Parquet files:

6.36 kB

Number of rows:

234