Computational Efficiency Assessment of Using Artificial Neural Networks in Structural Multiscale Finite Element Analysis

Multiscale modelling approach is commonly applied to accurately simulate mechanical behavior of structures made of heterogeneous materials, such as composites. For modeling heterogeneous materials using finite element method, homogenization process is performed where averaged material properties are determined from the microscale model. Then, these averaged properties are utilized in macroscale analysis, which allows to obtain proper displacement results under assumed load case. However, the stresses obtained from such macromodel do not describe the actual material behavior, and finding high-accuracy stress results leads to microscale computations for every integration point or node of the macromodel. Such operation is extremely time consuming therefore, the authors have proposed a method where the microstructural stresses are estimated on-the-fly by an artificial neural network from macroscale strain data. The following paper focuses on the measurement of accuracy and the increase in computational efficiency of implementing this method for an example composite structure.