Recent advances in machine learning open new opportunities to gain deeper insight into hydrological systems, where some relevant system quantities remain difficult to measure. We use deep learning methods trained on numerical simulations of the physical processes to explore the possibilities of closing the information gap of missing system quantities. As an illustrative example we study the estimation of velocity fields in numerical and laboratory experiments of density‐driven solute transport. Using high‐resolution observations of the solute concentration distribution, we demonstrate the capability of the method to structurally incorporate the representation of the physical processes. Velocity field estimation for synthetic data for both variable and uniform concentration boundary conditions showed equal results. This capability is remarkable because only the latter was employed for training the network. Applying the method to measured concentration distributions of density‐driven solute transport in a Hele‐Shaw cell makes the velocity field assessable in the experiment. This assessability of the velocity field even holds for regions with negligible solute concentration between the density fingers, where the velocity field is otherwise inaccessible.