The reverse design of a tunable terahertz metasurface antenna is proposed based on the deep neural network (DNN). To obtain the tunable properties of the terahertz antenna, the phase‐changed material vanadium dioxide (VO2) is introduced by controlling the voltage without changing the structures of the metasurface antenna. To improve the efficiency and accuracy of the terahertz antenna, the DNN is used to establish the relationship of amplitude and phase for the antenna unit. The prediction errors are below 10%. The radiation direction of the antenna can be changed from 0° to 52° by the VO2 with different states. The prediction error of angle is only 1° difference from the theoretical calculation. This reverse design method can be extended to other metasurface devices with improved efficiency and accuracy.