Viscosity is one of the important thermophysical properties of liquid aluminum alloys, which influences the characteristics of mold filling and solidification and thus the quality of castings. In this study, 315 sets of experimental viscosity data collected from the literatures were used to develop the viscosity prediction model. Back-propagation (BP) neural network method was adopted, with the melt temperature and mass contents of Al, Si, Fe, Cu, Mn, Mg and Zn solutes as the model input, and the viscosity value as the model output. To improve the model accuracy, the influence of different training algorithms and the number of hidden neurons was studied. The initial weight and bias values were also optimized using genetic algorithm, which considerably improve the model accuracy. The average relative error between the predicted and experimental data is less than 5%, confirming that the optimal model has high prediction accuracy and reliability. The predictions by our model for temperature- and solute content-dependent viscosity of pure Al and binary Al alloys are in very good agreement with the experimental results in the literature, indicating that the developed model has a good prediction accuracy.