Active optics is a key technology in the field of modern large reflective optical telescopes, which can effectively reduce the aberration and improve the imaging quality. The calibration algorithm depends heavily on the response matrix and physical parameters of the system. Due to the randomness and nonlinearity of the errors of the actual telescope system, the accurate response matrix and physical parameter model are often difficult to obtain, which leads to the unsatisfactory correction accuracy or the need for multiple corrections. To solve these problems, this paper proposes a deep learning calibration algorithm (DLCM) which does not depend on response matrix and physical parameter model. With the powerful prediction and self-learning ability of the deep neural network, this algorithm establishes the dynamic model network, strategy network, and decision-making unit needed by the correction algorithm. The control system can learn and optimize automatically by combining the corresponding equipment, so as to complete the mirror calibration work. Finally, using ANSYS finite element simulation to verify the DLCM algorithm, the results show that the proposed control algorithm can quickly and accurately complete the calibration work, and the calibration speed and accuracy are better than the traditional calibration algorithm.