Using recurrent neural networks for nonlinear component computation in advection-dominated reduced-order models