Due to its strong adaptability, deep learning has made significant contributions to solving engineering problems. This study proposed a temperature field prediction method for two-dimensional casting models based on deep learning. A convolutional neural network based on U-net was used to learn the variation law of the temperature field in casting models. Images representing the geometric shape of castings and thermal performance parameters were added to appropriate positions in the network, and temperature fields at different times with fixed time intervals were used as the training set. The trained model can predict the temperature field during the casting process and exhibits strong generalization in shape, material, and time, with an average prediction accuracy of 90%. Accuracy refers to the percentage of pixels where the absolute error value between the predicted and simulated results is less than 1% of the pouring temperature. At the same time, this network can make predictions quickly and has important applications in areas such as casting processes.