Abstract: An evaluation method based on adaptive radial basis function (RBF) neural network is proposed to solve the problems of low efficiency and poor analysis accuracy in the reliability assessment of thermal protection structures (TPS) under complex loads. The traditional grey wolf algorithm is improved by introducing a nonlinear convergence factor. The improved grey wolf algorithm is applied to optimize the number of central points and expansion constant of the radial basis function to establish an adaptive radial basis function neural network, which can accurately predict the stresses. The reliability of the thermal protection structure is numerically and experimentally investigated. It is concluded that the optimization performance of the grey wolf algorithm is significantly improved by introducing a nonlinear convergence factor. The proposed adaptive RBF model could quickly realize the data prediction through small samples while ensuring the accuracy. The reliability obtained by the method proposed matches well with that of Monte Carlo simulation and of experimental results.