GNNs Model Thermoplastic Composites, Accelerate Digital Twins

Short-fiber thermoplastic (SFT) composites are increasingly vital in lightweight structures, but their mechanical response is complex, governed by mesoscale interactions like fiber orientation, clustering, and porosity. Traditional finite element (FE) models, while capable of resolving this heterogeneity, are computationally prohibitive for realistic 3D microstructures. This research proposes a data-driven surrogate framework to efficiently predict the mechanical behavior of additively manufactured, compression-molded SFTs. Microstructures, reconstructed from micro-computed tomography data, are discretized into Voronoi-based cells, each representing distinct fiber-interaction neighborhoods. Nonlinear FE simulations homogenize each cell, generating stress-strain responses used to train a hybrid Graph Neural Network-Long Short-Term Memory (GNN-LSTM) architecture. The GNN-LSTM model effectively encodes microstructural topology and history-dependent mechanical evolution. It accurately predicts the stiffness and stress-strain behavior of unseen microstructures with an R² value of approximately 0.98, achieving a computational cost reduction of over two orders of magnitude compared to high-fidelity FE simulations. This approach provides a physics-informed, data-efficient pathway to identify mechanically weak microstructural cells and significantly accelerate the development of digital twins for SFT components.