Unidirectional fibre-reinforced polymers (UD-FRPs) are commonly used in a broad range of high-performance applications like aerospace, automotive or wind turbine rotor blades. The determination of their mechanical properties, for performance evaluation or as input for a more complex structural analysis, requires an excessive amount of testing due to their non-isotropic structure. Microscale finite element (FE) simulations using representative volume elements (RVEs) have shown to be able to predict the elastic material properties well, but require complicated algorithms to be set up and are computationally expensive and thus not suited for real-time applications.

This work presents a methodology to predict the elastic properties of UD-FRPs based on real micrographs combining the accuracy of microscale FE simulations with the efficiency and real-time capability of feedforward neural networks (FFNNs). A dataset for training, testing and validating the FFNN was created using microscale FE simulations with periodic boundary conditions. Therefore, multiple RVEs with varying spatial properties were generated using different fibre placement algorithms. The RVEs then served as a geometric baseline for carrying out FE simulations with varying material properties of the fibres and the matrix. The resulting material responses were homogenised to determine the elastic properties of the respective UD-FRPs as a function of the above-mentioned input properties. The resulting dataset was used to determine the structure of the FFNN using a Bayesian hyperparameter optimisation. The methodology itself was implemented in a software app and consists of three main steps: 1.) pre-process the micrograph to identify individual fibres and get their spatial properties using image processing, 2.) define the material properties of the constituents and a region of interest in the micrograph, 3.) predict the elastic properties of the region of interest using the trained FFNN. The real-time capability of the FFNN allows for the necessary responsiveness of the software app and thus enables a high evaluation efficiency.