The prediction of bedform configurations in rivers is very important for understanding the interaction among sediment-laden flows and hydraulic structures such as bridge piers and bridge abutments. The physics of bedforms is complex since it involves a broad range of spatial and temporal scales of turbulence; it also includes several phases (i.e., water, sediment,air) which are coupled. Further, bedforms are the result of the interaction among free surface waves and the riverbed from which particles are removed due to shear stresses. Often, the geometry of bedforms is predicted through diagrams, which contain nondimensional variables associated with the flow intensity, particlesize, and particle fall-velocity, among others. In this work, aneural network model is implemented to predict severalforms of bedform geometry. Five non-dimensional parameters are computed byusing the variables flow depth, flowrate, river width andslope, mean particle diameter and specific gravity of the sediment.The employed dataset includes 3006 laboratory data and 304field data, for which the bedform was known. Astratified cross validation approach was undertaken, and several optimization algorithms were assessed. It was found that bedforms can be reasonably well characterized by a neural network containing ten nodes in the intermediate layer. An overall performance of74.1% (1.9%) and an F1 score of 73.7% (2.1%) was obtained. Forall bedform classes, the prediction was higher than65%, except for the transition class, which is a bedform class considered a mixture of ripples and dunes. It was also found that the selection of the optimization algorithm does not significantly affect the overall performance of the neural network, although the Nadam algorithm provided a slightly better prediction compared to RMSprop and AdaMax algorithms.