Three different evaluation models on tribolocical performances of brake friction composites were established based on three types of typical artificial neural networks (ANN),including Elman,BP and RBF.All three models were trained and optimized with a Bayesian Regulation algorithm,and were applied to predict the friction coefficient of friction materials in both heating and cooling processes.The research results show that the Elman model is the best one in accurately predicting the friction coefficient of friction materials,especially for the formulations with a low usage of abrasives.