Stellar parameters and abundances for 3.2 million DESI DR1 spectra using machine learning

The work introduces a machine-learning pipeline for deriving stellar parameters and chemical abundances from spectra taken by the Dark Energy Spectroscopic Instrument (DESI). It builds on the model-driven Payne framework, training an artificial neural network on roughly 370,000 synthetic spectra of FGK stars. A key design choice is that the network fits each star's flux-calibrated, un-normalised spectral energy distribution (SED) directly, rather than a continuum-normalised spectrum as DESI's official Stellar Parameter (SP) pipeline does.

Internal accuracy tests show a median interpolation error below 0.3% for 90% of a synthetic verification sample, and the full fitting workflow recovers input labels with high accuracy even at low signal-to-noise ratio. External validation uses cross-matched samples of 6,719 APOGEE stars and 3,455 GALAH stars observed by DESI with S/N > 20; for these, the pipeline recovers effective temperature (Teff) and surface gravity (logg) with smaller systematic offsets than the DESI SP pipeline, an improvement attributed to using the full SED shape instead of a normalised spectrum.

The pipeline also recovers 12 elemental abundances—Na, Mg, Al, Si, Ca, Ti, V, Cr, Mn, Ni, Ba, and Y—at an accuracy that matches or exceeds SP where a comparison is possible. In addition, it provides four abundances that SP does not report at all: Mn, V, Ba, and Y. Applying the method to the DESI Data Release 1 stellar sample yields a catalogue of stellar parameters and abundances for 3.2 million stars.