Within turbulent boundary layers, the relationship between instantaneous surface momentum fluxes and streamwise velocities is more complicated than that between their ensemble averages described by the law of the wall. Although these fluxes need to be considered in large eddy simulations, the conventional approaches are not feasible. As an alternative, we have developed a deep neural network with the long short-term memory algorithmthat estimates instantaneous fluxes from a sequence of streamwise velocities. The velocities measured in a wind tunnel were used for training and validation. The trained deep neural network successfully estimates the instantaneous surface momentum flux with a suitable running average.