RESUMO
Local fiber alignment in fiber-reinforced thermoplastics is governed by complex flows during the molding process. As fiber-induced material anisotropy leads to non-homogeneous effective mechanical properties, accurate prediction of the final orientation state is critical for integrated structural simulations of these composites. In this work, a data-driven inverse modeling approach is proposed to improve the physics-based structural simulation of short glass fiber reinforced thermoplastics. The approach is divided into two steps: (1) optimization of the fiber orientation distribution (FOD) predicted by the Reduce Strain Closure (RSC) model, and (2) identification of the composite's mechanical properties used in the Ramberg-Osgood (RO) multiscale structural model. In both steps, the identification of the model's parameters was carried out using a Genetic Algorithm. Artificial Neural Networks were used as a machine learning-based surrogate model to approximate the simulation results locally and reduce the computational time. X-ray micro-computed tomography and tensile tests were used to acquire the FOD and mechanical data, respectively. The optimized parameters were then used to simulate a tensile test for a specimen injection molded in a dumbbell-shaped cavity selected as a case study for validation. The FOD prediction error was reduced by 51% using the RSC optimized coefficients if compared with the default coefficients of the RSC model. The proposed data-driven approach, which calculates both the RSC coefficients and the RO parameters by inverse modeling from experimental data, allowed improvement in the prediction accuracy by 43% for the elastic modulus and 59% for the tensile strength, compared with the non-optimized analysis.
