DOROTHY home

DOROTHY

Dorothy the dinosaur: a teal sauropod with gold star spots, standing on a stellar spectrum

Stellar labels from low-resolution spectra.

DOROTHY predicts the temperature, surface gravity and chemical abundances of a star from its DESI, BOSS or LAMOST spectrum. It learns from stars that APOGEE or GALAH also observed, and it reports an uncertainty for every label.

Install · Quick start · What DOROTHY does

Label transfer: a low-resolution spectrum of one star goes into a neural network; labels with uncertainties come out (Teff 4,812 ± 41 K, log g 2.61 ± 0.07, [Fe/H] −0.34 ± 0.04, [Mg/Fe] 0.12 ± 0.03) and place the star near the red clump of an illustrative Kiel diagram of about 650 stars coloured by [Fe/H]. Spectrum in one star, low resolution 4000 5000 6000 7000 8000 9000 wavelength (Å) flux Neural network trained on stars in common Labels out example Teff 4,812 ± 41 K log g 2.61 ± 0.07 [Fe/H] −0.34 ± 0.04 [Mg/Fe] 0.12 ± 0.03 The star among others Kiel diagram, illustrative 4000 6000 7000 0 1 2 3 4 5 Teff (K) log g main sequence turnoff subgiant branch red giant branch red clump 4812 2.61 [Fe/H] −1.5 +0.5
Label transfer: a low-resolution spectrum of one star goes into a neural network; labels with uncertainties come out (Teff 4,812 ± 41 K, log g 2.61 ± 0.07, [Fe/H] −0.34 ± 0.04, [Mg/Fe] 0.12 ± 0.03) and place the star near the red clump of an illustrative Kiel diagram of about 650 stars coloured by [Fe/H]. Spectrum in one star 4000 6000 8000 wavelength (Å) flux Neural network trained on stars in common Labels out example Teff 4,812 ± 41 K log g 2.61 ± 0.07 [Fe/H] −0.34 ± 0.04 [Mg/Fe] 0.12 ± 0.03 The star among others illustrative 4000 6000 7000 0 1 2 3 4 5 Teff (K) log g main sequence red giant branch red clump 4812 2.61 [Fe/H] −1.5 +0.5

Label transfer: a spectrum goes in, and labels with uncertainties come out; stars and values are illustrative.

Three steps

  1. Train on stars in common. Thousands of stars have both a low-resolution spectrum and high-resolution labels. A neural network learns to predict the labels from the spectrum.

  2. Predict for everyone else. The trained network predicts labels for the millions of stars that only the low-resolution survey observed.

  3. Report how sure it is. Every label comes with an uncertainty, calibrated on stars the network did not train on.

The data

All the data live in one HDF5 file, the super-catalogue, with one row per star, matched by Gaia DR3 source ID (Gaia Collaboration and Vallenari et al. 2023).

Name

Gives

Wavelengths

Resolving power

DESI

spectra

3600 to 9800 Å

2000 to 5000

BOSS

spectra

3600 to 10400 Å

1560 to 2650

LAMOST LRS

spectra

3700 to 9000 Å

≈ 1800

LAMOST MRS

spectra

4950 to 5350 Å, 6300 to 6800 Å

7500

APOGEE

labels

1.51 to 1.70 µm

≈ 22,500

GALAH

labels

four bands, 4713 to 7887 Å

≈ 28,000

The surveys are described in DESI Collaboration et al. (2022), DESI Collaboration et al. (2026) (DESI), Dawson et al. (2013), Smee et al. (2013) (BOSS), Cui et al. (2012), Liu et al. (2020), Zhao et al. (2012) (LAMOST), Majewski et al. (2017), Abdurro'uf et al. (2022) (APOGEE) and Buder et al. (2021), Buder et al. (2025) (GALAH).

Where to go next

  • Quick start: build a small synthetic catalogue, then train, evaluate and use a model on it.

  • The catalogue: the spectra and labels of every star, in one HDF5 file.

  • Train a model: write a configuration for your surveys and label sources, and train on it.

  • How it works: how the network learns each label and its uncertainty.