A method for controlling a deformable mirror surface shape based on a radial primary function processes samples of a neural network training set, obtains elements of the neural network training set, characterizes the complex surface shape of the deformable mirror in each sample by using a radial primary function, takes characterization parameters of the surface shape in the each sample as an input of a neural network, and trains the neural network by taking a voltage of each piezoelectric ceramic corresponding to the complex surface shape as a corresponding output of the neural network. Training times of the neural network are consistent with a number of the samples. Finally, the trained neural network is obtained to verify the training effect of the neural network. According to the characterization parameters of the required surface, the neural network is used to control the deformable mirror to generate the required surface.