In this research, the details of a study on using ANNs for prediction of damage severity in a model steel girder bridge are described. The dynamic tests conducted on the damaged and undamaged test structure showed that the reduction in stiffness during the damage lead to a reduction in natural frequencies for different modes. The experimentally obtained natural frequencies of the first five modes of the undamaged and damaged bridge model have been successfully applied as the training samples for the ANN. According to results in this study, ANN could predict the damage severity with an error of 5.6, 6.25 and 7.79% for training, testing and validation, respectively. The feasibility of ANNs as a powerful tool for predicting the severity of damage in a structure is investigated