CNN-based preform design: effect of training data configuration on strain distribution in forged products

This study investigates the effect of convolutional neural network (CNN)–based preform design on uniform strain distribution and microstructure in the forging of harmonic drive flex splines. Two CNN-based strategies were employed to achieve uniform strain distribution: one using NURBS (non-uniform rational B-splines) curves to generate a comprehensive training dataset and the other iteratively refining the training dataset by incorporating preform shapes from previous iterations to enhance learning. The CNN-designed preform reduced mean strain by 4.81%, standard deviation by 25.95%, and maximum strain by 19.22%. It eliminated folding defects and reduced forging load by 70.7% during preform forging and 7.40% during final forging. Electron backscatter diffraction (EBSD) confirmed the process’s effectiveness by comparing the microstructure of the forged flex spline. After isothermal normalizing heat treatment, the kernel average misorientation (KAM) value decreased by an average of 0.194° in forgings using existing preforms and by 0.249° with CNN-designed preforms.