Can exact conditions improve machine-learned density functionals?

Historical methods of functional development in density functional theory have often been guided by analytic conditions that constrain the exact functional one is trying to approximate. Recently, machine-learned functionals have been created by interpolating the results from a small number of exactly solved systems to unsolved systems that are similar in nature. For a simple one-dimensional system, using an exact condition, we find improvements in the learning curves of a machine learning approximation to the non-interacting kinetic energy functional. We also find that the significance of the improvement depends on the nature of the interpolation manifold of the machine-learned functional.