Surrogate constitutive model using artificial neural networks for the elastoplastic behavior of materials from fused filament fabrication

High-fidelity finite element simulations applied to materials from fused filament fabrication require high computational resources due to the non-linear behavior of the polymeric material and anisotropic characteristics induced by intrinsic heterogeneous morphologies observed at small scale. This work has the goal of developing a surrogate constitutive model based on artificial neural networks to enable faster finite element simulations of 3D printed structures. A multiscale approach is adopted to obtain the training database, on which a mesoscale representative volume element is defined based on experimental tests with printed samples. The representative element is submitted to different displacement and periodic boundary conditions in the finite element software, and effective stresses and strains components are obtained via periodic homogenization to integrate the database. Python programming language is employed to create the artificial neural network, and literature based hyperparameters are defined for predicting the material stress–strain relationships. The trained model is then coupled via user subroutine, and its validation is conducted by comparison with a representative volume element and experiments. It is observed that the surrogate method gave an adequate representation of the printed material mechanical behavior with a maximum difference in mechanical properties between numerical and experimental results reached 24.8%. Simultaneously, at a computational cost reduction of up to 56% for finite element simulations of such structures.