Prediction of shear angles in molten state thermo-stamped thermoplastic composites using a graph neural network
Please login to view abstract download link
Maintaining shear angles below a critical limit is crucial for the successful thermo-stamping of thermoplastic composites, helping to avoid wrinkles and inter-ply decohesion. This study introduces a data-driven approach, derived from nonlinear finite element simulations, to rapidly predict shear angles in thermo-stamped samples. The finite element model incorporates the necessary ingredients to capture the molten-state behavior of the composite employing a viscoelastic constitutive law fit to experimental data and proper contact loading. Each ply is represented as a superposition of membrane and plate finite elements, with material parameters that allow capturing the experimental behaviour. Additionally, a custom fiber frame is implemented to track fibers reorientation throughout the deformation process, ensuring an accurate description of the material anisotropy [1]. A graph neural network, MeshGraphNet (MGN) [2], is trained using data (finite element mesh, shear angles and displacements fields) generated by finite element simulations conducted on a parametric model, emphasizing the study of various mold geometries. Once trained, the MGN model predicts both shear angles and displacement fields for a given mold geometry. Two finite element datasets are employed, one comprising of a single ply composite deformed by a top (moving) and bottom mold, and a second involving a two-ply composite. As a proof-of-concept, the MGN model demonstrates promising performance in both prediction speed and accuracy relative to reference finite element simulations.
SPONSORS
Terms of service