By directly fitting the full shape of the spectra rather than a normalised spectrum, a machine learning model outperforms the official DESI pipeline and identifies four additional chemical elements.
Nagaraj Vernekar, Michael Hayden and Sara Lucatello have built an automatic learning analysis chain for the stellar spectra of the Dark Energy Spectroscopic Instrument (DESI). Inspired by The Payne, their network of neurons has been trained on approximately 370,000 synthetic spectra of F, G, and K type stars. Its peculiarity: it adjusts to the complete spectral energy distribution of each star, calibrated in flux, instead of a standardized spectrum as the official DESI pipeline does.
Validated on 6,719 stars already observed by APOGEE and 3,455 by GALAH, the method recovers effective temperature and surface gravity with lower systematic biases than the official pipeline. It also measures twelve chemical abundances, from sodium to yttrium, with at least equivalent accuracy, of which four (manganese, vanadium, barium and yttrium) were not supplied by DESI.
Applied to DESI's first publication of data, it produces a catalogue of parameters and abundances for 3.2 million stars, a precious material for galactic archaeology. The article was submitted to The Astrophysical Journal Supplement Series.
Read the full article on the original source
Go to the source →