Abstract: The neural network based on Bayesian method and its application to the mechanical property prediction were studied for hot-rolled SPA-H sheet. Integrated with the Occam's razor theory, a penalty term which could be interpreted as an indication of the complexity of the network was introduced into the objective function to prevent the occurrence of overfitting. Considering the practical problems in the application of the network and based on earlier work which have improved the Bayesian method, the Levenberg-Marquardt algorithm was employed to train the network, thus expediting the convergence rate. The network has been used in the prediction of mechanical properties of hot-rolled SPA-H sheet. Compared with the conventional BP neural network, it has the advantages of faster convergence rate, higher stability and ability for generalization.