2022MNRAS.516.3398N

While Gaia has observed the phase space coordinates of over a billion stars in the Galaxy, in the overwhelming majority of cases it has only obtained five of the six coordinates, the missing dimension being the radial (line-of-sight) velocity. Using a realistic mock data set, we show that Bayesian neural networks are highly capable of 'learning' these radial velocities as a function of the other five coordinates, and thus filling in the gaps. For a given star, the network outputs are not merely point predictions, but full posterior distributions encompassing the intrinsic scatter of the stellar phase space distribution, the observational uncertainties on the network inputs, and any 'epistemic' uncertainty stemming from our ignorance about the stellar phase space distribution. Applying this technique to the real Gaia data, we generate and publish a catalogue of posteriors (median width: 25 km s^–1) for the radial velocities of 16 million Gaia DR2/EDR3 stars in the magnitude range 6 < G < 14.5. Many of these gaps will be filled in very soon by Gaia DR3, which will serve to test our blind predictions. Thus, the primary use of our published catalogue will be to validate our method, justifying its future use in generating an updated catalogue of posteriors for radial velocities missing from Gaia DR3.