With the aperture of telescope becoming larger, the mass of primary mirror and other relevant structures will become heavier as well. Therefore, lighting weight for large space-based telescope is necessary. This paper purposed a method based on Neural Network aims to build a math model for primary mirror of large space-based telescope, which can reduce weight of the telescope and smaller mirror deformation caused by gravity release effectively. In the meantime, it can also improve stiffness of structure and reduce thermal strain caused by on orbit temperature variation effectively. The model describes the relationship between the structure of primary mirror of large space-based telescope and corresponding deformation, and describes the optical performance of mirror by using Zernike Polynomial. To optimize the structure of primary mirror lightweight, we take the deformation of mirror and its optical performance into consideration. To apply the structures parameters and its corresponding deformations to Neural Network training, we use the combination samples of different mirror lightweight structure parameters and corresponding deformation which caused by gravity release and thermal condition. Finally, by taking advantage of the Neural Network model to optimize the primary mirror lightweight of 1-meter rectangle space-based telescope, which can make the RMS 0.024λ (λ=632.8nm)and areal density under 15kg/m2. This method combines existing results and numerical simulation to establish numerical model based on Neural Network method. Research results can be applied to same processes of designing, analyzing, and processing of large space-based telescope directly.