Ilija Medan
The chemical properties of stars are normally derived from spectroscopic measurements, where such properties can give astronomers a glimpse into the environment in which these stars were born. However, this method can be difficult and more time consuming to obtain for low-mass stars in particular, due their relative faintness. On the other hand, photometry of low-mass stars is widely available due to the many all-sky surveys that have been conducted over the past couple of decades. To take advantage of this, we have calibrated a new relationship that predicts chemical properties of low-mass stars from multiple photometric measurements using machine learning with a Gaussian Process Regressor. We show that this regressor is able to predict these chemical properties with a high level of accuracy (+/-0.11 dex), while also mitigating systematic errors that were present in previous calibration attempts. Specifically, our technique avoids these systematic errors by removing unresolved binary star systems present in our sample before training the regressor, using an iterative method described here. The addition of this step in our method is crucial, as individual properties for the stars in unresolved systems cannot be derived accurately because of the blending of light from both objects. This newly calibrated relationship now allows for the chemical properties of ~10^7 stars in the vicinity of the Sun to be estimated. These metallicities can be utilized in many areas of research. For example, we discuss here how we plan to use this relationship to study local streams in the vicinity of the Sun, and determine how their kinematic properties are related to metallicity. Relationships between chemistry and kinematics of these groups will allow us to have a better understanding of the origins and histories of these streams, and relate them to dynamical interactions with spiral arms and/or bar-like features in our Galaxy.
To calibrate the relationship, we need a sample with the desired inputs and outputs; photometry and metallicity, respectively. The photometry comes from multiple surveys (Pan-STARRS, 2MASS and AllWISE) to provide us with measurements over a wide range in wavelength. From these surveys, all possible color combinations are used for the initial inputs, where “color” is the difference between two photometric measurements. Additionally, we obtain distance measurements from Gaia for all stars in our calibration sample, so we can calculate their absolute magnitudes (measure of the luminosity of stars), which are also used as initial inputs. For the outputs of our calibration sample, we use the average metallicity of the stars (the abundance of elements heavier than hydrogen or helium) from two spectroscopic surveys: APOGEE and Hejazi et al. Our final calibration sample comprises 6370 stars with required inputs (photometry) and outputs (metallicity).
To calibrate our relationship, we use the following procedure. For Step 1, we determine the optimal input colors and absolute magnitudes using an iterative method. This method first fits and evaluates the Gaussian Process regressor with all inputs. It then systematically removes each input, and fits and re-evaluates the regressor. The color/magnitude that has the least effect on the regression is removed. This process is repeated until only one input remains, and the optimal inputs are then those that minimized the mean squared error over the entire process. For step 2, we remove all unresolved binaries (UBs) from our calibration sample, as metallicity values derived from spectra of UBs are unreliable due to the blending of light from both stars in the system. This is done by first fitting 4th degree polynomials to HR diagrams (absolute magnitude vs. color) in bins of metallicity, [M/H]=0.1 dex. Stars that are over-/under-luminous compared to the fit, as expected of UBs, are removed. The fit is repeated and UBs removed until no more UBs are identified. When comparing the HR diagrams of the original sample (with UBs) and the cleaned sample (without UBs), the sample appears significantly cleaned. This is not apparent when comparing the samples on a color-color diagram. For Step 3, we redo Step 1 with the cleaned sample from Step 2, resulting in the final, optimal inputs for the relationship.
When examining photometric metallicity relationships from past studies, we find that there is a trend where metal-poor stars’ ([M/H]<0) metallicities are over-estimated and metal-rich stars’ ([M/H]>0) metallicities are under-estimated. We find that these systematic errors are present in our results when the regressor is trained with the original sample, but are largely absent when trained with the cleaned sample free of UBs.
With our improved relationship, we look at how the metallicity of local stars relates to their kinematics. We observe that there is a decrease in metallicity for stars with increased velocity. Broadly, this trend is due to the multiple components that make up the Milky Way. Additionally, we examine how specific components of velocity change with metallicity. Here we observe clumps in velocity space, whose structure is dependent on metallicity. With our relationship, we can study these changes in greater detail due to the much larger sample available from using low-mass stars. This will be helpful in understanding of the origins and histories of these streams, and how they tie into the dynamical history of the Milky Way.