In recent years, advances in deep learning techniques have led to the development of ‘normalising flows’, sophisticated algorithms for reconstructing arbitrary probability distributions given data sampled from them. In the context of galactic dynamics, with sufficient 6D stellar kinematics data, these techniques can be used to reconstruct the phase space distribution function (DF) in an entirely model-independent manner. This reconstructed DF can then be converted, via the collisonless Boltzmann equation, into an acceleration field.
In my talk, I will give a pedagogical demonstration of how (and how well) this technique works, stepping through a series of test cases of increasing sophistication. Then, I will turn to real data: the Gaia Catalogue of Nearby Stars (GCNS), and show the results of training normalising flows on the GCNS, which enables the reconstruction of the acceleration field in the Solar neighbourhood and in turn the local distribution of dynamical matter.