A bumper system plays a significant role in absorbing impact energy and buffering the impact force. Important performance measures of an automotive bumper system include the maximum intrusions, the maximum absorbed energy, and the peak impact force. Finite element analysis (FEA) of crashworthiness involve geometry-nonlinearity, material-nonlinearity, and contact-nonlinearity. The computational cost would be prohibitively expensive if structural optimization directly perform on these highly nonlinear FE models. Solving crashworthiness optimization problems based on a surrogate model would be a cost-effective way. This paper presents a design optimization of an automotive rear bumper system based on the test scenarios from the Research Council for Automobile Repairs (RCAR) of Europe. Three different mainstream surrogate models, Response Surface Method (RSM), Kriging method, and Artificial Neural Network (ANN) method were compared. First, design of experiment (DOE) was conducted to collect sample data which are used to build the three surrogate models. Accuracy of each surrogate model was then evaluated based on the root mean square error (RMSE), and the most effective surrogate model that approximates crashworthiness behavior of the rear bumper system was determined. Lastly, optimization was performed on the chosen surrogate model instead of the actual FE model to achieve an optimum lightweight design of the rear bumper system.