Prediction of aerodynamic heating under different flight conditions is a critical and challenging step in developing a new hypersonic vehicle. The prediction model usually involves a large number of variables, and this makes genetic programming converge too slow. This paper presents a fast mathematical modeling method, divide and conquer, for aerodynamic-heating predictions. It can use the separability feature of the target model to decompose a high-dimensional function into many low-dimensional sub-functions. The separability is detected by a special algorithm, bi-correlation test (BiCT), and the sub-functions could be determined by general genetic programming (GP) algorithms one by one. Thus the computational cost will be increased almost linearly with the increase of function dimension. This can help to break the curse of dimensionality and greatly improved the convergence speed to get the underlying target models from a set of sample data.