Evidence for a Vast Prograde Stellar Stream in the Solar Vicinity

TL;DR

The paper presents evidence for Nyx, a large stellar stream near the Sun, likely originating from a disrupted dwarf galaxy, revealed through Gaia data and deep learning analysis, indicating a past merger event in the Milky Way.

Contribution

This study identifies and characterizes Nyx, a new stellar stream, using Gaia data and deep learning, providing insights into the Milky Way's merger history and stellar dynamics.

Findings

Nyx is a coherent, prograde stellar stream near the Sun.

Nyx stars have distinct kinematics, lagging the disk by ~80 km/s.

Chemical and age data suggest Nyx originated from a dwarf galaxy.

Abstract

Massive dwarf galaxies that merge with the Milky Way on prograde orbits can be dragged into the disk plane before being completely disrupted. Such mergers can contribute to an accreted stellar disk and a dark matter disk. We present evidence for Nyx, a vast new stellar stream in the vicinity of the Sun, that may provide the first indication that such an event occurred in the Milky Way. We identify about 500 stars that have coherent radial and prograde motion in this stream using a catalog of accreted stars built by applying deep learning methods to the second Gaia data release. Nyx is concentrated within $\pm 2$ kpc of the Galactic midplane and spans the full radial range studied (6.5-9.5 kpc). The kinematics of Nyx stars are distinct from those of both the thin and thick disk. In particular, its rotational speed lags the disk by $\sim 80$ km/s and its stars follow more eccentric orbits. A small number of Nyx stars have chemical abundances or inferred ages; from these, we deduce that Nyx stars have a peak metallicity of [Fe/H] $\sim -0.5$ and ages $\sim $10-13 Gyr. Taken together with the kinematic observations, these results strongly favor the interpretation that Nyx is the remnant of a disrupted dwarf galaxy. To further justify this interpretation, we explicitly demonstrate that metal-rich, prograde streams like Nyx can be found in the disk plane of Milky Way-like galaxies using the FIRE hydrodynamic simulations. Future spectroscopic studies will be able to validate whether Nyx stars originate from a single progenitor.

Full text

In the $\Lambda$CDM paradigm, galaxies grow through the accretion of smaller satellites 1978MNRAS.183..341W . Accreted material from these mergers plays an important role in building the Milky Way’s stellar halo Johnston:1996sb , as demonstrated by the discovery of numerous tidal streams crisscrossing the sky (see 2016ASSL..420…87G for a review). In certain cases, an infalling satellite can interact dynamically with the stellar disk and be preferentially dragged into the Galactic plane 1986ApJ…309..472Q ; 1996ApJ…460..121W , leaving behind tidal debris that rotates with the disk. In this work, we present evidence for a new prograde stellar stream in the vicinity of the Sun, whose interpretation provides the first hint that such a merger occurred in our Galaxy.

A massive satellite falling towards the Milky Way’s disk experiences increased tidal friction that slows its orbit and eventually leads to its complete dissolution in the disk plane 1986ApJ…309..472Q . Feeling the pull of the satellite’s mass, disk stars lying ahead of the satellite lose angular momentum and fall into more central or “lower” orbits. In contrast, disk stars behind the satellite gain angular momentum and move into “higher” orbits. The net result is an overdensity of disk stars trailing behind the satellite that result in frictional drag. Such “disk dragging” is most pronounced for massive satellites ($\sim$ $1:10$ mergers) moving on prograde orbits. Observations of the Milky Way’s disk stars do allow for such an encounter 2008MNRAS.389.1041R ; 2009ApJ…691.1168H ; 2009MNRAS.397…44R ; 2009ApJ…703.2275P . The fact that accreted stellar and dark disks form from such mergers was first demonstrated using $N$-body simulations 1986ApJ…309..472Q ; 1996ApJ…460..121W ; 2008MNRAS.389.1041R ; 2009ApJ…703.2275P , and has since been reproduced in cosmological simulations with baryons 2009MNRAS.397…44R ; 2010JCAP…02..012L ; 2014ApJ…784..161P ; 2017MNRAS.472.3722G .

The primary difficulty in identifying accreted stellar debris near the Galactic midplane is distinguishing it from in situ stars born in the disk. This challenge is amplified when the tidal debris corotates with the disk. Indeed, standard selection methods for accreted stars conservatively ignore all stars with speeds $|v-v_{\text{lsr}}|\lesssim 220\,\text{km/s}$, where $v_{\text{lsr}}$ is the velocity of the local standard of rest. Chemo-dynamical template methods were used to search for an accreted stellar disk, but the analysis was not optimized for prograde structures with similar metallicities to the thick disk  Ruchti:2015bja . This motivates the use of techniques that can separate accreted stars from the in situ stellar background without significantly compromising statistics.

To address these issues, we used deep neural networks in Ostdiek et al. 2019 ml_paper to build a catalog of accreted stars from the second data release (DR2) of Gaia 2018arXiv180409365G . The network was trained on a combination of simulations and data. The first round of training used the m12i galaxy from the Ananke mock catalogs 2018arXiv180610564S , based on the Latte suite Wetzel2016 of Milky Way-like galaxies from the Fire simulations Hopkins:2013oba ; 2017arXiv170206148H . We used the star’s formation history as defined by Necib:2018igl to label a star as accreted if it was previously associated with a halo other than the host, and as in situ otherwise. Transfer learning was then performed with Milky Way stars, using the RAVE DR5–Gaia DR2 cross match 2017AJ….153…75K to label the most metal-poor stars ($\text{[Fe/H]}<-1.5$) at vertical distances $|z|>1.5\,\text{kpc}$ as accreted for the training.

The network takes as inputs the 5D kinematics of each star (two angular coordinates, two proper motions, and parallax) and then outputs a score associated with the probability that the star is accreted. When the network returns a score of $S=0$ ($S=1$), this implies that the algorithm has maximal confidence that the star is in situ (accreted). We assign all stars with $S>0.75$ to the catalog; this cut is intended to maximize the purity of the sample while minimally biasing the resulting kinematic distributions, and is justified by extensive testing performed on mock catalogs of simulated galaxies ml_paper . The network is optimized for and applied to the subset of Gaia DR2 stars with parallax $\varpi>0$ and fractional error $\delta\varpi/\varpi<0.10$.

In this work, we focus on the subset of Gaia DR2 stars with radial velocity measurements that fall in the region of interest (ROI) defined by spherical Galactocentric radii $r\in[6.5,9.5]\,\text{kpc}$ and vertical distances $|z|<3\,\text{kpc}$. We perform a Gaussian mixture analysis on the 3D velocities of all stars with a neural network score $S>0.95$. This sample consists of the stars that the network is most confident are accreted; focusing on this subset reduces the potential for disk contamination, although at the expense of possibly biasing the kinematic distributions ml_paper . A complete description of the mixture analysis, as well as other structures identified in the catalog, is provided in  Necib et al. 2019, hereafter Paper I catalog_paper . Here, we present evidence for a new stream discovered in this ROI, which comprises nearly 9% of the stars in the sample with $S>0.95$. We call this stream “Nyx,” after the Greek goddess of the night. We have verified that Nyx does not match any previously catalogued stream in 2010LNEA….4…13B ; 2016ASSL..420…..N ; 2018MNRAS.474.4112M ; 2018MNRAS.475.1537M ; 2018ApJ…860L..11K .

The left panel of Fig. 1 shows the Nyx stars in the Galactocentric spherical $v_{r}-v_{\phi}$ plane. The dashed gray lines denote the 1, 2, and 3$\sigma$ contours enclosing the stars with scores $S<0.75$. This subset is dominated by the in situ stars, but also has contributions from accreted stars that are more difficult for the network to classify. From Fig. 1, we see that the distribution is concentrated at $v_{\phi}\sim 220\,\text{km/s}$ and $v_{r}\sim 0\,\text{km/s}$. The 2 and 3$\sigma$ contours drop down to $v_{\phi}\sim 0\,\text{km/s}$ and non-zero $v_{r}$. This effect is likely due to stars that belong to accreted structures with highly radial velocities (see Paper I) that the network does not score as highly. Nyx stars lie at the tail of the thick-disk distribution, but extend much farther out into the kinematic region associated with the stellar halo.

The Nyx velocities are shown as arrows on a spatial projection of their locations in Fig. 1 (middle and right panels). In the following, we quote results in Galactocentric spherical coordinates. This choice is made to be consistent with the modeling of Gaia Enceladus in 2019ApJ…874….3N . Since this analysis is performed close to the disk plane, cylindrical and spherical coordinates are similar ($v_{R}\sim v_{r}$ and $v_{z}\sim v_{\theta}$). For Nyx, the means of the best-fit multivariate normal velocity distribution are $\{\overline{v}_{r},\overline{v}_{\phi},\overline{v}_{\theta}\}=\left\{134.0^{+3.1}_{-3.7},130.0^{+2.4}_{-2.3},53.0^{+3.5}_{-3.6}\right\}\,\text{km/s}$, with dispersions $\{\sigma_{r},\sigma_{\phi},\sigma_{\theta}\}=\left\{67.2^{+2.4}_{-2.5},45.8^{+1.8}_{-1.7},66.3^{+3.3}_{-3.0}\right\}\,\text{km/s}$. Nyx has a significant non-zero radial velocity that causes its stars to move at an angle in the $x-y$ plane. It is clearly prograde, moving with the Galactic disk, but lagging in velocity by $\sim$ 90  km/s. The 94 most likely stars to belong to Nyx are coherent in velocity, with total average speed 282 km/s and dispersion of 32 km/s. For reference, we provide a Toomre diagram of the Nyx stars in Supplementary Fig. 1. They are predominantly clustered within $\pm 1\,\text{kpc}$ of the midplane, passing through the Solar neighborhood, but they also extend up to $\pm 2\,\text{kpc}$. Nyx stars can be found through roughly the full radial range studied. From Fig. 1, it appears that we are missing stars at $x\sim-9$, $y<0\,\text{kpc}$ and $x\sim-7$, $y>0\,\text{kpc}$. We find that Nyx extends further in this direction when the neural network score cut is lowered to $S>0.85$ (see Supplementary Figs.  2 and 3). Lowering the cut to $S>0.85$ does not entirely fill in this region, however. It is likely that we are not capturing the entirety of the Nyx stream for a few reasons: the Gaia selection function, the network mis-labeling Nyx stars with low scores, and the use of a Gaussian clustering algorithm to identify stars belonging to (what is likely an) inherently non-Gaussian structure. Additionally, stars in this region have velocities pointing predominantly along the line-of-sight, a direction our machine learning algorithm is insensitive to having only been trained on parallax, positions, and proper motions ml_paper .

Few Nyx stars have accompanying spectroscopic measurements (see Supplementary Fig. 4). Cross-matching with RAVE-ON 2017ApJ…840…59C and applying a cut that the signal-to-noise ratio SNR $\geq$ 20, we find 12 stars that have measured [Fe/H] abundances, 3 of which also have [Mg/Fe] measurements. As shown in Fig. 2, the average value of [Fe/H] ([Mg/Fe]) is $-0.55^{+0.03}_{-0.04}$ ($0.18^{+0.10}_{-0.14}$), with respective dispersion of $0.13^{+0.03}_{-0.04}$ ($0.06^{+0.09}_{-0.05}$). The errors quoted here are obtained by generating 100 iterations of abundances of each star, sampling over the errors in each measurement, then calculating the means and dispersions. (The same figure using RAVE and GALAH data instead of RAVE-ON is shown as Supplementary Fig. 5.) For context, the average RAVE-ON measurement uncertainty on [Fe/H] ([Mg/Fe]) is 0.08 (0.11). For the small subset for which we have data, Nyx stars have abundances that are comparable with both the thick disk (see Fig. 2) as well as with dwarf galaxies 2004AJ….128.1177V ; 2009ARA&A..47..371T ; 2014MNRAS.444..515R . However, the small dispersion in the chemical abundances for Nyx does suggest a single progenitor origin, especially given the coherence in velocity space. Given the small subset of Nyx stars with abundances, and the large measurement uncertainties from RAVE-ON, further spectroscopic follow-ups (for example from APOGEE-2, 4MOST, and WEAVE) are needed to validate these conclusions.

In Fig. 3, we show the Nyx stars in the Hertzsprung–Russell (HR) diagram, in the Gaia bands $G_{BP}-G_{RP}$ and absolute magnitude $G+5\log(\varpi)+5$. We corrected for the reddening $E(BP-RP)$ and extinction $A_{G}$ using the associated Gaia values from 2018A&A…616A…8A . Due to possible contamination from the backgrounds, we choose stars with phot_bp_rp_excess_factor $<1.3+0.06~{}(G_{BP}-G_{RP})^{2}$, and only select low-extinction stars with $E(B-V)<0.015$ 2018A&A…616A..10G . Only 11 Nyx stars pass these photometric cuts. For reference, we overlay three isochrones with metallicities $\text{[Fe/H]}=-0.5$, -1, and stellar ages 1, 6, and 10 Gyr from the MIST project 2016ApJS..222….8D ; 2016ApJ…823..102C . The few Nyx stars that pass these photometric cuts are consistent with the older isochrones. This strengthens the case that Nyx originated from a disrupted satellite galaxy, because disk stars span a larger range of metallicities and ages and tend to be younger (see e.g. martig2016radial ).

We now turn to estimating the orbital parameters of Nyx stars. To do so, we use the gala package gala and assume the Milky Way potential of 2015ApJS..216…29B . Running the orbits of the stars backwards in time by 1 Gyr, we find that the mean eccentricity is $0.60$ (dispersion of $0.12$), mean pericenter is $3.9\,\text{kpc}$, mean apocenter is $16\,\text{kpc}$, and mean $z_{\text{max}}$ is $1.7\,\text{kpc}$ (see Supplementary Figs. 6 and 7). These orbital properties are distinct from what is expected for the thin disk, which has eccentricities close to circular. The orbits are also more eccentric than expected of the thick disk, which has eccentricities peaked at $e\sim 0.3$ and concentrated below $e\sim 0.5$ 2017ApJ…850…25L . In Supplementary Fig. 7, we show the orbits traced by the Nyx stars integrated back by 1 Gyr; the prograde and highly eccentric motion of the Nyx stars is evident. Coupling this observation with the fact that Nyx lags behind the disk by $\sim 90\,\text{km/s}$ and has a significant radial velocity component makes a strong case that it is the result of a satellite merger.

Stellar “moving groups” can arise from resonances in stellar orbits that are perturbed by the Galactic bar Dehnen_2000 ; Fux:2001sm or density waves caused by the spiral arms 2004MNRAS.350..627D ; Quillen:2005uk . However, simulations of bar and spiral arm perturbations typically result in coherent velocity structures that lag the Galactic disk by only $\sim 20$–$50\,\text{km/s}$ 2010LNEA….4…13B . It is in principle conceivable that a stronger perturbation to the disk could have been indirectly induced by the passage of a satellite galaxy through the disk, without necessarily leaving many stars behind: the dynamical “kick” caused by such a passage can result in density waves moving in the same direction as the Sun at velocities comparable to what we find for Nyx 2009MNRAS.396L..56M ; 2009MNRAS.397.1599Q . However, such perturbations generically result in a specific shape (“arches”) in the cylindrical $v_{R}-v_{\phi}$ plane. This effect has been observed in Gaia DR2 2018A&A…616A..11G , and does not resemble the kinematics of Nyx. Indeed, as will be discussed later, there are hints for additional members of the Nyx stream that may fall along constant eccentricity surfaces Helmi:2005wc , as expected of a satellite merger, rather than constant energy surfaces, as expected for a disk perturbation 2009MNRAS.397.1599Q .

Taking this evidence together, we argue that Nyx is a remnant of a dwarf galaxy. Its characteristics are largely consistent with previous studies of accreted stellar disks 2008MNRAS.389.1041R ; 2009MNRAS.397…44R ; 2009ApJ…703.2275P ; 2017MNRAS.472.3722G . Namely, it is corotating with a lag speed that falls in the range motivated by simulations. Additionally, its mean metallicity of $\text{[Fe/H]}~{}\sim~{}-0.55$ is consistent with a galaxy of stellar mass in the range $[4\times 10^{9},8\times 10^{9}]\,M_{\odot}$, as estimated using the mass-metallicity relation of 2013ApJ…779..102K . Note that we account for both the theory and RAVE-ON measurement uncertainties in this estimate. The range quoted is bracketed by the 16 and 84 percentiles. The lower end of this range is compatible with the mass scale of the satellites studied in 2008MNRAS.389.1041R ; 2009MNRAS.397…44R ; 2009ApJ…703.2275P ; 2017MNRAS.472.3722G that led to viable accreted stellar and dark disks. In general, disk dragging of satellites is most efficient for such massive systems. However, we do caution that this is a simple estimate, and does not account for the potential metallicity gradients that would be expected as various layers of stars are tidally stripped from the satellite progenitor. Dedicated simulations of different merging scenarios will establish the feasibility of Nyx, either as an accreted satellite or a thick disk perturbation.

To demonstrate that such metal-rich prograde streams can exist in galaxies like our own — and that our analysis procedure correctly identifies them if they do — we repeat our study on the mock catalog of the Fire simulation m12f, focusing on the region within $2.3\,\text{kpc}$ of LSR1 2018arXiv180610564S . The general properties and merger history of the m12f galaxy are described in 2018ApJ…869…12S ; Necib:2018igl . Proceeding in a parallel manner to our Gaia study, we train the network on the m12i galaxy, then apply transfer learning on m12f stars. It is important to note that m12i and m12f are completely different (but both Milky Way-mass) galaxies from different regions in a cosmological volume. Their merger histories, in particular, are quite distinctive: the evolution of m12f is more active up to redshift $\sim 1$, while m12i is more quiescent at late redshifts 2017arXiv170206148H .

To identify individual populations in velocity space, we run the Extreme Deconvolution mixture analysis on the subset of m12f stars with network scores $S>0.95$ using four Gaussian distributions. Note that this is different from the method used to identify Nyx (see Paper I). It is computationally faster, but does not yield uncertainties on the best-fit parameters, which we do not need for this case. Two Gaussians (shown in blue and red on the left panel of Fig. 4) are matched with the largest stream of m12f, which was labeled as Merger I in Necib:2018igl . We refer to these as SI and SII, respectively. The progenitor is a massive satellite that merged at redshift $\sim 0.12$–0.39, with a peak dark matter halo mass of $1.5\times 10^{11}\,M_{\odot}$, and a total stellar mass right before merger of $1.8\times 10^{9}\,M_{\odot}$. Its tidal debris has an average stellar metallicity of $\text{[Fe/H]}\sim-0.88$ Necib:2018igl , alpha abundance \text{[\alpha/Fe]}=0.26, and stellar age of $9.6$ Gyr.

The right two panels of Fig. 4 show the spatial distribution of the velocities for SI and SII as red and blue arrows, respectively. Similarly to Nyx, both streams corotate with the disk and are concentrated within a few kpc of the midplane. Both SI and SII are highly radial, but move in opposite directions towards/away from the galactic center. This suggests that they correspond to two passages of the stream near the solar position in m12f. The total velocity of SI (SII) is 440 (419) km/s with a dispersion of 29 (11) km/s. Future studies of the simulations will unveil the different possible stream velocities and dispersions expected from such mergers, which will enable us to better understand the origin of Nyx.

We can also check if there are other stars in the $S>0.95$ sample that originated from the same progenitor as SI and SII, but that are not identified by the mixture analysis. These stars are shown in pale orange in the left panel of Fig. 4. Notably, SI and SII comprise only a fraction of the stream in the $S>0.95$ sample, likely because the mixture analysis separates the individual populations assuming Gaussian velocity distributions, and the stream distribution is inherently non-Gaussian. SI and SII comprise an even smaller fraction of the total stream, because demanding a score cut of $S>0.95$ significantly reduces the number of stream stars in the sample; only $19\%$ of the m12f stream stars pass this cut. This fraction increases to 41% when reducing the cut on the network score to $S>0.85$, though at the expense of introducing more disk contamination to the sample. Looking at the purity of the system (defined as the fraction of stars passing the score cut that are truly accreted), we find that it is $53\%$ for $S>0.95$ but drops to 43% for $S>0.85$.

The example from the m12f simulated galaxy serves as a proof-of-principle that not only a prograde stream can comprise a non-trivial fraction of the accreted stars near the Solar neighborhood, but also that accreted stars from the disk can be identified using this method. This test case also suggests that additional remnants of the Nyx stream may exist that are not captured in the $S>0.95$ sample, or by the mixture model analysis. To test the first hypothesis, we repeat our study using the $S>0.85$ sample of accreted stars (see Paper I for details). We find a population of stars with properties consistent with Nyx, as well as an additional stream (referred to as Nyx-2) with similar prograde motion and chemical abundances, but opposite radial velocity. This suggests that Nyx-2 is related to Nyx and might actually be debris from a separate passage of the same satellite, similar to the m12f example above (see Supplementary Fig. 2). Dedicated simulations will be important in unveiling the origin of Nyx, to identify the mass and infall time of the progenitor if it is a merging object, and its subsequent effect on the disk.

Follow-up spectroscopic studies will play a crucial role in establishing the origin of the Nyx stream by providing a larger set of chemical abundances. This will help either confirm the kinematic arguments above that suggest that Nyx is the remnant of a dwarf galaxy, or that it is a perturbation of the Milky Way disk. Additionally, if Nyx-2 and Nyx have similar elemental abundances, it would substantiate their relation to each other. Our study of a simulated Milky Way-like galaxy demonstrates that infalling dwarf galaxies can leave signatures very similar to what we observe for Nyx. If Nyx is indeed the result of such a merger, then it would provide evidence for accreted prograde stars, and potentially, an accompanying dark matter component Necib:2018igl in a stream or disk. The presence of such a dark matter component would substantially alter our current understanding of the local dark matter phase-space distribution, and have important ramifications for terrestrial searches for the dark matter particle.

Data Availability

The accreted star catalog used for this analysis is available at https://doi.org/10.5281/zenodo.3579379. The simulation m12f is available at https://fire.northwestern.edu/ananke/. The IDs of Nyx stars are available as part of the Supplemental Material of this work.

Code Availability

This analysis makes use of emcee and Extreme Deconvolution for the Gaussian Mixture model. The Python Markov chain Monte Carlo code emcee is freely available and documented at http://dfm.io/emcee/current/. Extreme Deconvolution is freely available at https://github.com/jobovy/extreme-deconvolution. Details regarding the application of these two public codes are provided in catalog_paper .

Author Contribution

All authors discussed the results and commented on the manuscript. ML and TC conceived the project. LN built the data analysis pipeline. BO, TC, and MF conceptualized the machine learning algorithms. BO built the deep neural network and produced the accreted stellar catalog. Interpretation of the results and writing of the original manuscript were done by LN and ML. The FIRE-2 simulation code was built by PFH, and run by PFH, SGK, and AW. SGK and AW ran the halo finding algorithm on the simulation, and LN identified the mergers as a function of redshift. RS built the mock catalogs used in the training of the neural network.

Corresponding Author

Correspondence to Lina Necib (lnecib@caltech.edu).

Acknowledgments

We thank A. Helmi, J. Johnson, E. Kirby, N. Laracy, J. Read, N. Shipp, and J. Wojno for helpful discussions. This work was performed in part at Aspen Center for Physics, which is supported by National Science Foundation grant PHY-1607611. This research was supported by the Munich Institute for Astro- and Particle Physics (MIAPP) of the DFG cluster of excellence “Origin and Structure of the Universe.” This research was supported in part by the National Science Foundation under Grant No. NSF PHY-1748958.

LN is supported by the DOE under Award Number DESC0011632, and the Sherman Fairchild fellowship. ML is supported by the DOE under contract DESC0007968 and the Cottrell Scholar Program through the Research Corporation for Science Advancement. BO and TC are supported by the US Department of Energy under grant number DESC0011640. MF is supported by the Zuckerman STEM Leadership Program and in part by the DOE under grant number DE-SC0011640. SGK and PFH are supported by an Alfred P. Sloan Research Fellowship, NSF Collaborative Research Grant #1715847 and CAREER grant #1455342, and NASA grants NNX15AT06G, JPL 1589742, 17-ATP17-0214. AW is supported by NASA, through ATP grant 80NSSC18K1097 and HST grants GO-14734 and AR-15057 from STScI, and a Hellman Fellowship from UC Davis. This work utilized the University of Oregon Talapas high performance computing cluster. Numerical simulations were run on the Caltech compute cluster “Wheeler,” allocations from XSEDE TG-AST130039 and PRAC NSF.1713353 supported by the NSF, and NASA HEC SMD-16-7592.

RES thanks Nick Carriero, Ian Fisk, and Dylan Simon of the Scientific Computing Core at the Flatiron Institute for their support of the infrastructure housing the synthetic surveys and simulations used for this work.

This work has made use of data from the European Space Agency (ESA) mission Gaia (http://www.cosmos.esa.int/gaia), processed by the Gaia Data Processing and Analysis Consortium (DPAC, http://www.cosmos.esa.int/web/gaia/dpac/consortium). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the Gaia Multilateral Agreement.

Funding for RAVE has been provided by: the Australian Astronomical Observatory; the Leibniz-Institut fuer Astrophysik Potsdam (AIP); the Australian National University; the Australian Research Council; the French National Research Agency; the German Research Foundation (SPP 1177 and SFB 881); the European Research Council (ERC-StG 240271 Galactica); the Istituto Nazionale di Astrofisica at Padova; The Johns Hopkins University; the National Science Foundation of the USA (AST-0908326); the W. M. Keck foundation; the Macquarie University; the Netherlands Research School for Astronomy; the Natural Sciences and Engineering Research Council of Canada; the Slovenian Research Agency; the Swiss National Science Foundation; the Science $\&$ Technology Facilities Council of the UK; Opticon; Strasbourg Observatory; and the Universities of Groningen, Heidelberg and Sydney. The RAVE web site is at https://www.rave-survey.org.

I Methods

A detailed presentation of the analysis methods is provided in two companion papers. For a discussion of the machine learning methods, including critical validation tests, see ml_paper . For a discussion of the clustering methods and further interpretation of the results when applied to the accreted star catalog, see catalog_paper .

Appendix A Supplemental Material

This analysis makes use of emcee 2013PASP..125..306F and Extreme Deconvolution 2011AnApS…5.1657B for the Gaussian Mixture model.