Stochastic Trust-Region Response-Surface Method (STRONG) is an improved response surface method that possesses automation property and convergence guarantee. One important assumption underlying STRONG is that the response surface is of the same form as the metamodels, i.e., either a first- or second-order model. In many real problems, however, the response surface is very nonlinear and unpredictable. In this paper, we investigate the optimal first- and second-order designs in the presence of model misspecifications. We propose a sequential design framework that integrates these designs to improve the computational efficiency of STRONG. Numerical experiments verify the effectiveness of the proposed sequential design framework.