In adaptive optics systems, the traditional proportional-integral control model relies on the response matrix of the deformable mirror, which is sensitive to changes in the system state. When the response matrix is altered, the wavefront correction performance is degraded. In this paper, the output of control signal from Hartman slope data is realized by redefining the back-propagation neural network structure, and a control model is established. Experimental results show that the proposed model eliminates the limitation of the traditional fixed model and acquires the characteristics of an online real-time update response model. The control model delivers high convergence performance, can adapt to environmental changes, and is robust. It also improves the control precision and the control performance to a certain extent.