Dedulek/gaia-eb-teff-datasets

Gaia Eclipsing Binary Effective Temperature Datasets

Multi-survey photometry and derived (T_{\mathrm{eff}}) products for Gaia DR3 eclipsing binaries. Used by gabdeevmaksim/gaia_eb_rf_teff to train and deploy a soft PCA–KDE ensemble of four Random Forest regressors.

Summary (paper production)

Quantity	Value
Cross-matched catalogue	(2,169,721)
Valid GSP-Phot	(1,276,166) (58.8%)
Soft-KDE ML fills (no GSP-Phot)	847 486
Final Teff coverage	97.9%
Still without Teff	46 069 (2.1%)

Files

photometry/eb_unified_photometry.parquet

Primary input for training and prediction (Gaia + auxiliary photometry, (E(B-V)), EB parameters, GSP-Phot labels where available). Missing-value sentinel: -999.0.

python scripts/download_datasets.py --datasets photometry
# → data/raw/eb_unified_photometry.parquet

correction/teff_correction_coeffs_deg2.pkl

Degree-2 polynomial + Hermite blend ((T_0=10,000,\mathrm{K}), (W=2000,\mathrm{K})) applied to hot GSP-Phot training labels.

python scripts/download_datasets.py --datasets correction
# → data/teff_correction_coeffs_deg2.pkl

predictions/eb_catalog_teff.parquet

Merged catalogue product (GSP-Phot original/corrected + ML / ensemble columns) produced by scripts/merge_teff_predictions_into_unified.py. Soft-KDE no-GSP-Phot fills are also available as ensemble_teff_no_gspphot_predictions.parquet from a local ensemble run.

Soft PCA–KDE (not Best-of-Three)

Deployed temperatures for objects without GSP-Phot use soft mixture weights from local good/bad KDEs in a shared 2D PCA plane, with EB-params fallback out of coverage. Holdout: MedAE 311 K, MAE 675 K, 68.6% within 10%. Ignore older Hub text quoting 263 K / 97.2% Best-of-Three metrics.

Citation

Cite the accompanying paper and the code repository. Gaia products remain under ESA archive terms.