Multiscale Thermodynamics-Informed Neural Networks (MuTINN) for nonlinear computations of CNT-coated fuzzy fiber composites accounting for viscoplasticity and interfacial damage.

Overview & Synopsis

Fuzzy fiber is a special type of reinforcement architecture, in which a regular fiber has carbon nanotubes (CNTs) grown on its surface, through chemical vapor deposition improving hence the interphase. To efficiently model these composites, an equivalent fiber approach is adopted combining several homogenization techniques. This simplification enables systematic parametric analyses, though it remains computationally demanding. Moreover, effectively implementing these methodologies in real scale applications, particularly within FE 2 analyses, remains challenging due to their substantial computational cost. To circumvent this issue, the present study proposes a data-driven alternative based on Artificial Neural Networks. The approach yields a macroscopic surrogate model for CNT-coated fuzzy fiber composites, designed to account for geometric complexity and nonlinear material response. This framework, called Multiscale Thermodynamics Informed Neural Network (MuTINN), integrates thermodynamic principles in a specially designed two ANNs system. The MuTINN captures the evolution of Internal State Variables and Helmholtz free energy, thereby eliminating the need for memory-based network architectures. Offline data is supplied by finite element simulations performed on representative volume element (RVE) subjected to complex loading conditions. Trained on these datasets, the MuTINN achieves accurate predictions of stress, strain, and energy, incorporating the composite's inelastic behavior and heterogeneous microstructure. Moreover, to test the predictive capabilities of MuTINN, the model was validated against finite element simulations, including loading paths not used during model training, demonstrating the model's performance. In addition, the influence of dataset preparation strategies was investigated, showing that random data splitting across time-steps improves the consistency between training, validation, and testing results.