Approximating FE Draping Simulations using AI-based Surrogates

The development of manufacturing processes is commonly assisted by numerical simulations using the finite element (FE) method. However, especially in the preforming phase of composites, slack material properties make the simulation process quite time- and computationally expensive. Especially in recent years, research is therefore trending towards surrogate models using artificial intelligence (AI)-based methods to approximate the behavior and accelerate the simulation process. In literature, a great variety of AI-based approaches can be found in various implementations ranging from simple approaches such as linear regression to sophisticated programs using neural networks and deep-learning based surrogates [1 – 4].
The aim of this research is to explore the potential of surrogate modeling for the approximation of FE draping simulations. Therefore, we created an FE simulation model using the commercial software package Abaqus/CAE, representing the draping process of a glass fiber fabric onto a rib tool. The model is parametrized to allow for an automated FE training data generation. Fig. 1 shows the FE draping simulation model including the draping tool and the blank cut. A detailed description of the model setup and the custom routine for automatic training data generation can be found in [5]. For the approximation of the FE simulation, we analyzed and compared two approaches of different complexity: an image-based surrogate model and an X-position based dynamics (XPBD) approach. In case of the image-based surrogate model, we created training data by modifying the pressure magnitudes on three equal sized areas on the top surface of the blank cut surface. We trained a U-Net using the input parameters and their corresponding output encoded as greyscale images. For the second approach using the XPBD surrogate, the FE model serves as a baseline enabling the evaluation of the draping result. To reduce the complexity of the draping task, the blank cut was replaced by a square blank cut for the XPBD approach. The XPBD simulation model is combined with a reinforcement learning (RL) agent, which determines the optimum draping path parameters to smooth out a wrinkle on the blank cut (see Fig. 2). To approximate the material behavior, simplified material properties were implemented and adjusted using the baseline FE model.
Comparisons between FE and surrogate simulation showed that both approaches enable a precise prediction for the topology of the draped result (see Fig. 3). For the image-based surrogate, differences between simulation and surrogate only occur at fringes of draping defects (see yellow areas in Fig. 3a). It is a simple implementation that provides fast predictions; however, it is not geometry invariant, meaning that different draping geometries require new training. The XPBD also shows high compliance with the FE simulation within the area of interest (see Fig. 3b). It allows for interchangeable geometries and thus provides a generalizable approach. Both approaches provide promising results for elevating the technology readiness level (TRL) up to the use in modern composite manufacturing.