GSPPhot-metallicity calibration - Gaia
Gaia DR3 GSP-Phot metallicity calibration
As discussed, e.g. in Andrae et al. 2022, the [M/H] estimates provided by GSP-Phot in Gaia DR3 exhibit strong systematic errors. As a result, they should not be used for quantitative analysis without a suitable calibration. Here, one possible calibration is briefly described, but the user is explicitely encouraged to develop their own calibration.
The objective is to use a multivariate adaptive regression spline (Friedman 1991, hereafter MARS) in order to learn a mapping from GSP-Phot's biased [M/H] to some well-established metallicity estimates. Various literature catalogues were considered as possible training samples and eventually LAMOST DR6 was opted for because it provides a broad range of metallicity values but does not probe too deeply into high-extinction regions in the Galactic disk. Since LAMOST provides [Fe/H] estimates, the MARS model not only needs to remove the systematics from GSP-Phot's [M/H] but also translate from [M/H] to [Fe/H]. Given the metallicity bias in GSP-Phot also depends on stellar parameters, the input features of the MARS model include the effective temperature, surface gravity, the biased [M/H] value itself and the extinction and reddening. It also includes Galactic latitude, which helps with the translation from [M/H] to [Fe/H]. The trained MARS model then provides the calibrated [Fe/H].
The Python package is initialised and called as follows:
The input provided to the tool is a pandas data frame which must contain the following columns (using Gaia DR3 column names):
- positions either as ra,dec or l,b
Obviously, the metallicity calibration tool is not perfect. Its task is to improve the (otherwise hardly usable) [M/H] estimates from GSP-Phot. The community is explicitely invited to develop better calibration tools. Here, we list several limitations:
- Calibrations are only given for the MARCS library (Teff from 2500K to 8000K) and the PHOENIX library (Teff from 3000K to 10000K) but, due to a lack of training data, not for the A and OB libraries.
- The metallicity calibration works very well on low-extinction stars but not so well on high-extinction stars. For example, the metallicity differences to GALAH DR3 values (low-extinction sample) are reduced by ~30% for MARCS and ~50% for PHOENIX. Conversely, for APOGEE DR16 (high-extinction sample), the metallicity differences are reduced by only ~10% for PHOENIX and actually slightly increased for MARCS.
- The calibration should not be used outside the training sample range of LAMOST DR6 (Teff from ~3800K to ~8500K, [Fe/H] from -2.5 to +1).
- Since we train on LAMOST DR6 estimates of [Fe/H], any systematic errors in LAMOST DR6 are inherited.