Polymer extrusion die design using a data-driven autoencoders technique

Designing extrusion dies remains a tricky issue when considering polymers. In fact, polymers exhibit strong non-Newtonian rheology that manifest in noticeable viscoelastic behaviors as well as significant normal stress differences. As a consequence, when they are pushed through a die, an important die-swelling is observed, and consequently the final geometry of the extruded profile differs significantly from the one of the die. This behavior turns the die’s design into a difficult task, and its geometry must be defined in such a way that the extruded profile results in the targeted one. Numerical simulation was identified as a natural way for building and solving the inverse problem of defining the die, leading to the targeted extruded geometry. However, state-of-the-art rheological models reveal inaccuracies for the desired level of precision. In this paper, we propose a data-driven approach that, based on the accumulated experience on the extruded profiles for different dies, learns the relation enabling efficient die design.