A volume-driven CNN framework replacing expert intuition in 3D forging preform design

초록

This study proposes a convolutional neural network (CNN)-based 3D preform design framework to prevent defects such as underfill and folding in hot forging processes while minimizing forging load and flash formation. We converted simulation data into voxels for training and constructed initial datasets using Laplace-based isosurface geometries along with cuboid and cylindrical shapes. To enhance generalization capability, deformation-based data augmentation was employed. For the quantitative evaluation of forged products regarding preform geometry, the proposed forging volume efficiency index (FVEI) integrates both geometric conformity and critical defect indicators into a single, comprehensive metric. Preforms designed for three representative geometries successfully reduced forging load and flash while eliminating defects. The proposed approach demonstrated reliable performance even for previously unseen geometries that were not included in the training set. This framework presents a fully automated design methodology independent of expert knowledge, highlighting its strong potential for direct application in industrial settings. © 2025

키워드