• 목록

[Abstract]

This paper details the development of a convolutional neural network-based surrogate model for the finite element method (FEM) to iteratively predict material behavior during H-beam rolling. The inherent complexity and multi-stage nature of the H-beam rolling process render conventional FEM analysis computationally prohibitive. As a solution, we leveraged 3D FEM simulations to create a comprehensive training dataset. Nodal data from these simulations was systematically converted into images using Delaunay triangulation. The resulting model can predict key physical variables—material shape, temperature, equivalent strain (EQ strain), equivalent stress (EQ-stress), and normalized Latham-Cockcroft damage—at the conclusion of each rolling pass. Process conditions were integrated as a mask layer to reduce error accumulation during iterative predictions, thereby enhancing overall accuracy. The U-Net model, enhanced with partial convolution that operates exclusively on valid pixels within the image, demonstrated a 0.0567 higher shape F1-score compared to a standard U-Net. We evaluated the prediction performance in a multi-stage rolling process which is a difficult process for prediction because of error accumulation. Despite the difficulty, the surrogate model achieved reasonably good performance, a shape F1-score of 0.9871 and drastically reduced the prediction time from 32 hours for a traditional FEM simulation to just 0.85 seconds. Furthermore, the model attained a Mean Absolute Percentage Error (MAPE) of less than 22% for the iterative prediction of four material properties, indicating an error level lower than anticipated.