In finite element analysis, increasing mesh density significantly enhances the accuracy of simulation results, but it also demands a greater consumption of computational resources. In order to solve this contradiction, Res2Net, U-Net, channel attention mechanism and geometric feature extraction were integrated to learn the cloud image data of finite element results with low mesh density. Predict high mesh density finite element resulting cloud images, thereby reducing the computational costs required without sacrificing accuracy. By conducting experiments under different scale conditions such as 2 ×, 4 × and 8 ×, the mean square error and mean absolute error of the test data have been significantly reduced, which fully proves the excellent performance of the model in the accuracy of numerical prediction. The results indicate that with minimal computational resource investment and guaranteed high precision of output results, this model can be effectively utilized for the upscaling of finite element result cloud images.