Machine learning approach for the prediction of forming results of fiber-reinforced thermoplastic composites using simulation and curvature data

Forming of fiber-reinforced thermoplastics is a promising process for the mass production of lightweight components. However, due to the complex material behavior, various types of defects may occur during forming, which affects component performance. The formation of wrinkles is the primary deformation phenomenon investigated in computational costly numerical forming simulations, typically by analyzing the shear angle distribution. Therefore, this contribution presents a machine learning-based approach using a convolutional autoencoder to predict the shear angle distribution of forming geometries. The approach leverages the curvature distribution of the geometry as input for the shear angle prediction, resulting in significantly reducing computation time compared to high-fidelity simulations. A database is developed based on the four principal curvatures (planar, parabolical, elliptical, hyperbolical) commonly found in forming geometries. Hyperparameter studies and the influence of database composition are investigated. Validation is conducted using an unseen, complex double dome geometry, followed by tests on additional geometries to evaluate model generalization. Results show that shear angle distributions can be accurately predicted using a simple curvature geometry database. The trained model thus enables rapid evaluation of forming behavior, accelerating the development process for fiber-reinforced thermoplastic components in thermoforming applications.