CREATING TRAINING DATA FOR SURROGATE MODELS USING FE DRAPING SIMULATION

In the last years, Finite Element (FE) simulations became increasingly important for the development and optimization of manufacturing processes. They enable a reduction of trial-and-error experiments in the development phase, saving costs and time. Despite the benefits, FE simulations, especially for fiber reinforced composites, can take a considerable amount of time to compute. A promising approach to address this issue is the approximation of the process behavior, using computationally inexpensive, artificial-intelligence-based surrogate models. When training such models with sufficient training data, they can predict the outcome for a given input accurately. However, a common issue for this approach is the lack of training data. Therefore, this work focuses on the efficient generation of training data for a surrogate model, using FE analysis. Towards this goal, we developed an automated routine for data generation. A Python script generates input files, post-processes the output files and exports relevant data. The underlying parametrized draping simulation model is called with different input parameters, to simulate various draping cases. This approach can generate training data sets, covering around 200 different parameter variations, within reasonable time. To evaluate the suitability of the FE results as training data, we trained a U-Net based surrogate model. The resulting surrogate model predictions were compared with the respective simulation results. Additionally, to determine the minimum amount of training data necessary for the surrogate model to predict results correctly, we conducted a parametric study with varying training set sizes.