Estimates of the atmospheric parameters of M-type stars: a machine-learning perspective

Abstract

Estimating the atmospheric parameters of M-type stars has been a difficult task due to the lack of simple diagnostics in the stellar spectra. We aim at uncovering good sets of predictive features of stellar atmospheric parameters (T eff , log (g), [M/H]) in spectra of M-type stars. We define two types of potential features (equivalent widths and integrated flux ratios) able to explain the atmospheric physical parameters. We search the space of feature sets using a genetic algorithm that evaluates solutions by their prediction performance in the framework of the BT-Settl library of stellar spectra. Thereafter, we construct eight regression models using different machine-learning techniques and compare their performances with those obtained using the classical χ 2 approach and independent component analysis (ICA) coefficients. Finally, we validate the various alternatives using two sets of real spectra from the NASA Infrared Telescope Facility (IRTF) and Dwarf Archives collections. We find that the crossvalidation errors are poor measures of the performance of regression models in the context of physical parameter prediction in M-type stars. For R ∼ 2000 spectra with signal-to-noise ratios typical of the IRTF and Dwarf Archives, feature selection with genetic algorithms or alternative techniques produces only marginal advantages with respect to representation spaces that are unconstrained in wavelength (full spectrum or ICA). We make available the atmospheric parameters for the two collections of observed spectra as online material.

Machine Learning Applied to Star–Galaxy–QSO Classification and Stellar Effective Temperature Regression

“…The machine learning has been adopted as a successful alternative approach to defining reliable objects classes, stellar types and types of variable stars (eg. Liu et al 2015;Kovács & Szapudi 2015;Krakowski et al 2016;Kuntzer et al 2016;Sarro et al 2018;Pashchenko et al 2018). It is not the first time to take advantage of this technology to classify the objects or to regress the stellar parameters.…”

“…Liu et al (2015)'s research also implies that a large sample could cover a larger area of the parameter space, and further could yield more re-liable prediction. Sarro et al (2018) constructed regression models to predict T eff of M stars with eight machine-learning algorithms. The training sample is built with the features extracted from the BT-Settl of synthetic spectra.…”

“…Then, the models were applied to two sets of real spectra from the NASA Infrared Telescope Facility and Dwarf Archives collections. Sarro et al (2018) used the root mean/median square errors (RMSE/RMDSE) to describe the prediction errors. The RMSEs were from 160 to 390 K, and the RMDSEs from 90 to 220 K, various with the different algorithms and signal-to-noise ratios.…”

Surveying the reach and maturity of machine learning and artificial intelligence in astronomy

“…Other techniques that have been investigated, often in conjunction with one or more the above methods, include: AdaBoost (Bethapudi & Desai, 2018;Xin et al, 2017); genetic algorithms (Sarro, Ordieres-Meré, Bello-García, González-Marcos, & Solano, 2018); SOMs (Armstrong et al, 2018;Armstrong, Pollacco, & Santerne, 2017;Süveges et al, 2017); recurrent neural networks (Naul et al, 2018); autoencoders (Sedaghat & Mahabal, 2018;Vincent, Larochelle, Bengio, & Manzagol, 2008); and transfer learning (Benavente, Protopapas, & Pichara, 2017). Falling within the generation and reconstruction category (Section 2.2), GANs are likely to be the next most significant machine learning approach for astronomy.…”

The CARMENES search for exoplanets around M dwarfs

“…They report that their convolutional neural network achieved better accuracy than the other ML algorithms, and point out the importance of a sufficiently large training set. Sarro et al (2018) used genetic algorithms for selecting features such as equivalent widths and integrated flux ratios from BT-Settl model atmospheres. They estimated T eff , log g, and [M/H] for M dwarfs with eight different regression models and ML techniques, and compared the results to classical χ 2 and independent component analysis coefficients.…”