Multiscale Thermodynamics-Informed Neural Networks (MuTINN) for nonlinear structural computations of recycled thermoplastic composites

Abstract

Fiber-reinforced thermoplastic composites are increasingly valued for their lightweight properties, mechanical performance, and recyclability, yet the recycling process introduces microstructural heterogeneities that degrade their mechanical behavior. To address the challenges from a modeling point of view, this study proposes a Multiscale Thermodynamics-Informed Neural Network (MuTINN) approach to predict the nonlinear, anisotropic response of recycled glass fiber-reinforced polyamide 6 composites, with the primary aim of enabling structural simulations in significantly reduced time compared to traditional FE${}^2$ approaches. The MuTINN framework integrates thermodynamic principles with artificial neural networks (ANNs) to capture the evolution of internal state variables and Helmholtz free energy, eliminating the need for memory-based networks. Finite element simulations of representative volume elements (RVEs) under diverse loading conditions are utilized to provide off-line data for the MuTINN. The latter accurately predicts stress, strain, and energy quantities, accounting for the anisotropic and heterogeneous nature of recycled materials. While trained using numerical simulations at 0° and 90° orientation specimens, the proposed framework succesfully predicts the response for specimens with 45° orientation with error in the maximum stress level up to 1.6%. The model is implemented into commercial finite element analysis (FEA) software via a Meta-UMAT framework, allowing efficient macroscale simulations. Validation against experimental data and finite element-based periodic homogenization confirms the framework’s accuracy for structural computations.