The Survey of Surveys - machine learning for stellar parametrization

Aug 4, 2026

Upload by Asher

Machine learning is used to infer stellar parameters—effective temperature, surface gravity, and iron metallicity—from photometric data of large sky surveys such as SDSS and SkyMapper. The approach leverages a previously built, homogenized and recalibrated spectroscopic catalog (from surveys like APOGEE, GALAH, and LAMOST) to train a neural network. Results reach spectroscopic-quality for millions of stars observed only photometrically, with typical uncertainties around 100 K in temperature, 0.1 dex in log g, and 0.1 dex in [Fe/H], including challenging low-metallicity regimes.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

AccessFree to download

Introduction

Input Catalogs and Training Set

| Publication Year | 2024 | | --- | --- | | Acceptance in OA@INAF | 2025-02-07T16:41:01Z | | Title | The Survey of Surveys: machine learning for stellar parametrization | | Authors | TURCHI, Alessio; PANCINO, Elena; ROSSI, Fabio; AVDEEVA, Aleksandra; MARRESE, Paola Maria; et al. | | DOI | 10.1117/12.3018967 | | Handle | [http://hdl.handle.net/20.500.12386/35860](http://hdl.handle.net/20.500.12386/35860) | | Journal | PROCEEDINGS OF SPIE | | Number | 13101 | The Survey of Surveys: machine learning for stellar parametrization Turchi, A.a , Pancino, E.a,b , Rossi, F.a , Avdeeva, A.c , Marrese, P.d,b , Marinoni, S.d,b , Sanna, N.a , Tsantaki, M.a , and Fanari, G.b aINAF-Osservatorio Astrofisico di Arcetri, L.go Enrico Fermi 5, Firenze, Italy b Space Science Data Center, Via del Politecnico SNC, I-00133 Rome, Italy c Institute of Astronomy, Russian Academy of Sciences, 48 Pyatnitskaya St. , Moscow 119017, Russia dINAF – Osservatorio Astronomico di Roma, Via Frascati 33, 00040, Monte Porzio Catone, Roma, Italy ABSTRACT We present a machine learning method to assign stellar parameters (temperature, surface gravity, metallicity) to the photometric data of large photometric surveys such as SDSS and SKYMAPPER. The method makes use of our previous effort in homogenizing and recalibrating spectroscopic data from surveys like APOGEE, GALAH, or LAMOST into a single catalog, which is used to inform a neural network. We obtain spectroscopic-quality parameters for millions of stars that have only been observed photometrically. The typical uncertainties are of the order of 100K in temperature, 0.1 dex in surface gravity, and 0.1 dex in metallicity and the method performs well down to low metallicity, were obtaining reliable results is known to be difficult. Keywords: machine learning, big data, surveys, stars, spectroscopy 1. INTRODUCTION In the last few years large spectroscopic surveys provided a huge amount of photometric measurements for hundreds of millions (up to billions) of stars, either at low-medium resolution such as the RAdial Velocity Experiment (RAVE) [1], the Sloan Extension for Galactic Understanding and Exploration (SEGUE) [2] and the Large sky Area Multi Object fiber Spectroscopic Telescope (LAMOST) [3], or at high-resolution such as the Galactic Archaeology with HERMES (GALAH) [4], the Apache Point Observatory Galactic Evolution Experiment (APOGEE) [5] and the Gaia-ESO survey [6] . The data provided by these surveys allow to give a precise estimates of key parameters such as effective temperature (Teff), surface gravity (log g) and iron metallicity ([Fe/H]) for few millions of stars in the Milky Way. The availability of high-quality spectroscopic measurement, together with a good estimation of distance and reddening, is of paramount importance to derive high quality estimates of the above parameters from photometric surveys. In recent years, these surveys spawned a many works focused on Machine Learning (ML) methods (i.e. Neural Networks or simpler methods), such as the Cannon [7], the Payne [8] and StarNet [9] . ML saw a huge development in the last decades of XX century and rose to a widespread usage in the first decades of XXI century. The term can be used as a general hat to cover different disciplines from Artificial Intelligence to Neural Networksand Computational Statistics. In general we refer to ML techniques when based on algorithms that make use of heterogeneous data to automatically “learn” and build a “model” that is used to produce a desired output, using statistical methods. ML methods applied to spectroscopic star catalogues mainly focus on the analysis of the provided data to many different purposes, i.e. trying to enhance the physical model used to compute the parameters, thus this field is already mature in the case where the previous parameters are directly derived from spectroscopic data. However spectroscopic quality measurements are hard to perform on large stellar samples,

FAQs

What stellar parameters does the method predict from photometric data?

It predicts effective temperature, surface gravity, and iron metallicity ([Fe/H]).

How is training data prepared for the neural network?

Spectroscopic-quality labels are taken from the SoS catalogue, built by homogenizing and recalibrating spectroscopic measurements from multiple surveys.

What accuracy is reported for the inferred parameters?

Typical uncertainties are about 100 K for temperature, 0.1 dex for surface gravity, and 0.1 dex for metallicity, with good performance at low metallicity.

Aug 4, 2026

Upload by Asher