Neural networks with dimensionality reduction for efficient springback prediction in deep drawing of multi-material cylindrical cups

Abstract

ABSTRACT Deep drawing involves the use of a punch to plastically deform a workpiece to create sheet metal parts such as cups and channels. Upon release of the punch, the workpiece undergoes elastic recovery and results in a change in geometry known as the springback effect. The process is highly non-linear and involves many parameters, leading to large computation times to fully simulate the process or train a regression model on the high dimensionality of the data. In this paper, dimension-reduced neural networks (DR-NNs) for efficient springback prediction in the process of deep drawing a cylindrical cup are presented with consideration for different materials. The DR-NNs are trained on FEM data from a deep drawing simulation performed on ABAQUS/CAE. Features such as the workpiece’s initial coordinates, material properties, and thickness are introduced as inputs for the dimensionality reduction. Linear and non-linear dimensionality reduction methods reduce the input to a smaller set of principal components, which are fed as inputs to the neural networks for predicting the springback after punch release. The DR-NNs are compared against a deep neural network (DNN) and show improvements in lower computation time for training, prediction uncertainty, and less storage space required while retaining prediction accuracy.