ApogeeNet - SDSS

On this page

Summary
Detailed Description
Data Products
Validation and Uncertainties
Flags

Summary

ApogeeNet (Olney et al. 2020; Sprague et al. 2022; Sizemore et al. 2024 ) uses a convolutional neural network to estimate stellar labels (effective temperature, surface gravity, and metallicity) given a rest-frame resampled APOGEE spectrum. The network is trained on high quality APOGEE spectra and labels from SDSS-IV Data Release 17.

Detailed Description

ApogeeNet predicts the stellar parameters given a rest-frame APOGEE spectrum. Flux values smaller than 10^-6 are clipped, then log-transformed. We then compute the 95th percentile of the log-transformed flux values, and clip any log flux values higher than this value. This preprocessing is largely to treat imperfect sky subtraction.

Flux arrays are randomly resampled based on the uncertainties, and each sample is used to make predictions of stellar parameters. We compute summary statistics (median and standard deviation) from those distributions of stellar parameters and report those as the catalog values.

The summary ApogeeNet files released for Data Release 19 are based on APOGEENet v3. There are two versions of ApogeeNet integrated in Astra that are relevant for Data Release 19: APOGEENet versions 2 and 3. ApogeeNet v2 was used to provide initial guesses for ASPCAP, but are not formally released in Data Release 19. In ApogeeNet v2 the stellar parameters were predicted given an APOGEE spectrum, 2MASS photometry, and Gaia astrometry. In ApogeeNet v3, the stellar parameters are predicted only given an APOGEE spectrum (i.e., no photometry).

Data Products

In Data Release 19, ApogeeNet is executed on the combined spectra stored in `apStar` data products, and the rest-frame visit spectra that are also stored in `apStar` files. This includes SDSS-V APOGEE spectra reduced with version 1.4 of the APOGEE data reduction pipeline, and SDSS-IV APOGEE spectra reduced for Data Release 17.

There are two key data products produced by APOGEENet and Astra:

`astraAllStarAPOGEENet`: this file contains stellar parameters based on combined spectra in the `apStar` files
`astraAllVisitAPOGEENet`: this file contains stellar parameters based on visit spectra in the `apStar` files

Validation and Uncertainties

Using repeat observations of the same stars we computed the z-score distribution from all random pair-wise matches, and found that the raw uncertainties reported by ApogeeNet were slightly underestimated. For these reasons, we inflate the ApogeeNet stellar parameter uncertainties with the following relations:

$\sigma_{\rm Teff} = 1.25 \sigma{\rm Teff,raw} + 10$
$\sigma_{\rm logg} = 1.25 \sigma_{\rm logg,raw} + 0.01$
$\sigma_{\rm [Fe/H]} = \sigma_{\rm [Fe/H],raw} + 0.01$

The ApogeeNet result files include both the raw uncertainties reported by ApogeeNet (noted by the prefix `raw_`), and the recommended values.

Figure 1: Figure from Olney et al. 2020 showing the distribution of T_eff and log g values for the YSOs across the various regions. The compilation of the measurements from the IN-SYNC pipeline (top panel), photometrically derived input labels (middle panel), and the resulting output from the APOGEE Net (bottom panel) are shown. The gray lines show the isochrones at ages of 1, 2, 5, 10, 20, 50, 100, 200, and 300 Myr, and 8.5 dex, as well as the evolutionary tracks for 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1 solar mass stars from the PARSEC isochrones (Marigo et al. 2017).

Olney et al. 2020 showed the improvement from APOGEENet in the parameters of pre-main-sequence stars through comparison of the Kiel diagram for star-forming region. Figure 1 shows a figure from their paper demonstrating that ApogeeNet parameters have a distribution that is expected from their photometric distribution. ApogeeNet was trained on labels derived from photometry for these stars, leading to more physical parameters. In contrast the parameters from the IN-SYNC pipeline (Cottaar et al. 2015), which were derived by fitting to a set of BT-SETTl synthetic spectra, have log g’s that are systematically too high.

Flags

Field Name	Bit	Flag Name	Flag Description
`result_flags`	0	`TEFF_UNRELIABLE`	Effective temperature is unreliable because it is outside the range 3.1 < log10(TEFF) < 4.7
`result_flags`	1	`LOGG_UNRELIABLE`	Surface gravity is unreliable because it is outside the range -1.5 < logg < 6
`result_flags`	2	`FE_H_UNRELIABLE`	Metallicity is unreliable because it is outside the range -2 < [Fe/H] < 0.5 or log10(TEFF) > 3.82
`result_flags`	3	`E_TEFF_UNRELIABLE`	Error on effective temperature is unreliable
`result_flags`	4	`E_LOGG_UNRELIABLE`	Error on surface gravity is unreliable
`result_flags`	5	`E_FE_H_UNRELIABLE`	Error on metallicity is unreliabl
`result_flags`	6	`E_TEFF_LARGE`	Error on effective temperature is large
`result_flags`	7	`E_LOGG_LARGE`	Error on surface gravity is large
`result_flags`	8	`E_FE_H_LARGE`	Error on metallicity is large
`result_flags`	9	`MISSING_PHOTOMETRY`	Missing photometry
`result_flags`	10	`RESULT_UNRELIABLE`	Stellar parameters are knowingly unreliable
`result_flags`	11	`NO_RESULT`	Exception raised when loading spectra
`flag_warn`	—	`WARN`	If E_TEFF_LARGE or E_LOGG_LARGE or E_FE_H_LARGE or MISSING_PHOTOMETRY is set.
`flag_bad`	—	`BAD`	If any of the following are set: RESULT_UNRELIABLE, TEFF_UNRELIABLE, LOGG_UNRELIABLE, FE_H_UNRELIABLE, E_TEFF_UNRELIABLE, E_LOGG_UNRELIABLE, E_FE_H_UNRELIABLE, or NO_RESULT.