Speaker
Description
The increasing electrification of the mobility sector leads to some additional weight due to the implementation of heavy batteries and associated peripherals. Therefore, lightweight materials such as fiber-reinforced thermoplastics (FRTPs) are required to increase efficiency through weight reduction without compromising the mechanical properties of the component. Most automotive FRTP lightweight parts are produced by the thermoforming process, which has the greatest potential for mass production of flat-shelled components such as the underbody protection for the battery in battery electric vehicles (BEV). However, due to the complex draping behaviour of the material, various defects can occur during the forming process, such as wrinkling due to shear deformation, which has a negative impact on the properties of the part. Wrinkles lead to localised deflections in the flow of forces, which have a negative effect on the structural-mechanical properties. In addition, they can be the starting point for crack initiation in structurally relevant areas of the part, so appropriate part and process design is essential to avoid critical wrinkle formation. Such a design of the forming process is usually modelled as experimental testing is too expensive. However, this is computationally expensive due to the complex material behaviour. Therefore, this paper deals with the implementation of surrogate models for predicting the forming result, which are more time-efficient than numerical simulations. The surrogate model used is a convolutional autoencoder that predicts the shear angle distribution as an indicator of possible wrinkling. These convolutional neural networks (CNNs) are well suited to classification and pattern recognition tasks on data with a known grid-like topology, such as images. The input is the curvature distribution of randomly shaped geometries. A study of selected hyperparameters is carried out to obtain an appropriate shear angle prediction. The model is trained with shear angle distribution results from finite element simulations, as these can physically accurately represent the complex material behaviour. Therefore, a workflow has been implemented in Python that generates equidistant meshes of random geometries, automatically transfers them to a numerical simulation, and performs data generation that serves as training and test data sets for the convolutional autoencoder. Following hyperparameter tuning, this paper analyses whether the surrogate model trained with random curvature distributions is also able to predict the shear angle distribution of real forming geometries such as elliptical structures.
