Neural network reads a million star spectra to measure temperature and chemistry

The work

Technical · from the record

A deep convolutional neural network built on the astroNN package was trained to read LAMOST low-resolution optical spectra and return effective temperature, surface gravity, [M/H], [Fe/H], [α/M] and nine individual elemental abundances, using calibrated APOGEE DR17 measurements of stars observed by both surveys as training labels. On the test set the mean absolute error was 29 K for Teff, 0.07 dex for log g, 0.03 dex for [Fe/H] and [M/H] and 0.02 dex for [α/M], with most elements between 0.02 and 0.04 dex. The trained network was then run on 1,210,145 LAMOST DR8 giants, including sub-giants, and the resulting catalogue was compared with GALAH DR3 measurements for 11,949 stars in common, where the residual bias in Teff was about 25 K against the 20 K bias between APOGEE and GALAH themselves. The authors report that abundances of metal-poor stars and of manganese, nickel and calcium below [Fe/H] = -1.5 dex are supported by few or no training samples and should be treated carefully.

How AI was used

LAMOST DR8 spectra with g-band signal-to-noise above 10 were cross-matched against APOGEE DR17 stars with complete Teff, log g, [Fe/H], [M/H] and [α/M], restricted to 3500-5500 K and log g 0.0-4.0 dex. Spectra were truncated to 4000-8500 Å, interpolated to one flux value per angstrom and continuum-normalised by Gaussian smoothing with a 50 Å kernel using the dataset module from The Cannon; spectra without LASP redshifts or heavily affected by cosmic rays were removed. The remaining stars were split randomly into reference and test sets in a 9:1 ratio, and the reference set further into training and cross-validation sets at 9:1. Labels exceeding per-element uncertainty limits or lying outside subjectively chosen effective parameter ranges were set to NaN rather than removing the star, since the astroNN custom loss function ignores missing labels and down-weights high-uncertainty ones. A weight matrix prepended to the loss raised the weight of metal-poor stars, capped at 10, and set the weight of stars with -0.5 < [Fe/H] < 0.1 dex to 0.95. The model was a convolutional network from the astroNN class BayesianCNNBase with three convolutional layers, a max-pooling layer after the third, two fully connected layers, and MCDropout layers for dropout variational inference; uncertainties combined model uncertainty from Monte Carlo dropout forward passes with predictive uncertainty. All labels were learned and predicted simultaneously. The trained network was applied to the LAMOST DR8 giant sample, and additional per-element sub-networks, each predicting one abundance along with Teff, log g and [Fe/H], were trained as a supplement.