The accuracy of airfoil polar predictions is limited by the usage of imperfect turbulence models. Can machine-learning improve this situation? Will airfoil polars teach the effect of turbulence on skin-friction? We try to answer these questions by refining turbulence treatment in the Rfoil code: boundary layer closure relations are learned from airfoil polar data. Two turbulent closure relations, for skin friction and energy shape factor, are parametrized with a class-shape transformation. An experimental database is then used to define code inaccuracy measures that are minimized with an interior point gradient algorithm. Results show that airfoil polars contain exploitable information about turbulent phenomena. Inferred closures agree with direct numerical simulation results of skin friction and the new code predicts drag more accurately. Maximum lift remains under-predicted but Rfoil maintains its robustness and suitability for optimization of wind energy airfoils.