Combining FE and Multiscale Thermodynamics Informed Neural Networks towards Fast and Frugal Inelastic Simulations for Woven Composite Structures
Please login to view abstract download link
Achieving accurate predictions for inelastic woven composites, while considering their microstructures, appears feasible through the application of multi-scale modeling approaches. Incorporating these methodologies into real-scale applications, particularly within FE² analyses, remains challenging due to their significant computational requirements [1]. To overcome this issue, while considering the scale effects, this study presents a novel approach utilizing Artificial Neural Networks (ANNs) to create a macroscopic surrogate model for composites. Termed as Multiscale Thermodynamics Informed Neural Networks (MuTINN) [2], this model is founded on thermodynamics laws and introduces specific quantities of interest as internal state variables at the macroscopic level. Effectively capturing the state and evolution laws governing the history-dependent behavior of composites, MuTINN ensures both thermodynamic admissibility and the physical interpretability of overall responses. To facilitate composite structure computations, a FE-MuTINN approach has been successfully developed to integrate MuTINN into a finite element (FE) code, utilizing a Meta-UMat. The efficiency of this approach is demonstrated across various material scales, with specific examples involving woven composite structures. These applications account for anisotropic yarn damage and elastoplastic polymer matrix behavior. The numerical results, alongside comparisons to experimental data and FE simulations, demonstrate the prediction capabilities of the MuTINN approach. It also reveals the high level of consistency across a wide spectrum of complex non-proportional loading conditions.
