Implicit neural representations for deep drawing and joining experiments

A deep understanding of metal deformation processes is essential for producing complex
geometries in many industrial applications. Although simulations using Finite Element
Methods (FEM) have helped in steering toward that goal, they are particularly time-consuming for large 3D meshes. Searching for the process parameters that lead to the desired shape of a metal part can become extremely expensive in terms of man-hours and computational resources. We investigated how machine learning models, especially deep  neural networks, can help in speeding up the design process of deep drawing and joining processes by allowing a fast interpolation of FEM simulations from minutes or hours to seconds. In this study, inspired by implicit representations of 3D objects using neural networks, an implicit approach is used to predict local properties such as the thickness of the metal sheet, its thinning, and plastic strain, using solely the process parameters defining the experiment. We observe that the low number of trainable parameters of the predicting model ensures a generalization to unseen process parameters and ultimately allows for a reliable fast inspection of the processes.