CupNet -- Pruning a network for geometric data

The optimization of production processes can benefit from machine learning methods that incorporate domain knowledge and data from numerical simulations [1]. Typically, such methods aim to model relations between process parameters and the resulting product. In this manuscript, we consider an example from the field of deep drawing, a sheet metal forming process in which a sheet metal blank is drawn into a forming die by mechanical action. Specifically, we study the prediction of product geometries in a cup drawing process based on data from finite element simulations [2].