The paper presents an algorithm combining square response surface and Radial Basis Function Neural Network (RBFNN) to solve the problem that RBFNN is often difficult to meet the precision request of approximation model. By using this method the extended error of the approximation model is diminished, the approximation precision is improved, and the flexibility is enhanced based on having the same sample points. Two numerical examples indicate that the method is effective in increasing the approximation precision and can be used to increase the design efficiency and quality in multidisciplinary design optimization (MDO).