The surface accuracy of the large-aperture reflector antenna has a significant influence on the observation efficiency. Recent researchers have focused on using the finite element (FE) simulation to study the effect of gravity and heat on the deformation distribution of the main reflector. However, the temperature distribution of the antenna is challenging to obtain, and it takes a long time for the FE simulation to carry out FE modeling and post-processing. To address these limitations, this study presents a surrogate model based on Extreme Gradient Boosting (XGBoost) and deep Convolutional Neural Network (CNN) to get the deformation distribution of the main reflector quickly. In the design of the surrogate model, using the XGBoost algorithm and sparse sampling to solve the difficulty of obtaining the entire temperature distribution is first proposed, and then a deep CNN is developed for estimating deformation. Based on the effect of dynamic loads on the antenna structure, a diverse dataset is generated to train and test the surrogate model. The results show that the surrogate model reduces the calculating time dramatically and can obtain the indistinguishable deformation compared to the FE simulation. This technique provides a valuable tool for temperature and deformation calculation of large-aperture antennas.