Zeiteffiziente und datenfreie Bauteil und Prozesssimulation mithilfe von Physics Informed Neural Networks

Tobias Würth, along with Anabel Prietze, Clemens Zimmerling, Constantin Krauß, and Luise Kärger from the Karlsruhe Institute of Technology (KIT), presented a novel approach to component and process simulation. The focus was on leveraging Physics-Informed Neural Networks (PINNs) for efficient, data-free simulations in lightweight construction. The primary challenge addressed in the presentation was the time and resource-intensive nature of numerical simulations (like FEM) and physical experiments for complex product optimization. Traditional efficiency measures, such as data-driven surrogate models (e.g., neural networks), often require extensive data generation and have limited reliability outside the data range. The team at KIT introduced an innovative idea: integrating known physics into a surrogate model instead of relying solely on observational learning. This approach involves using PINNs, which incorporate physics (PDEs) and allow for data-free training, integrating existing data to produce parametrized, grid-free solutions. The methodology hinges on Physics-Informed Neural Networks, enabling data-driven simulations. A significant highlight was the ability of PINNs to perform data-free training using differential equations and feed-forward neural networks. This method allows for rapid quantitative outputs and qualitative support through grid-free visualizations, significantly enhancing product and process understanding. The application of this method was demonstrated in the context of a Carbon Fiber Reinforced Polymer (CFRP) plate in an autoclave process. The team showcased how PINNs could be used for material, geometry, and process adjustments, providing visual aids for engineers to understand the impacts of these adjustments. The PINN approach supports both qualitative and quantitative aspects of design decisions. It enables the output of solutions within fractions of a second on standard desktop PCs, with a relative mean deviation from FEM of about 0.1%. The method also offers qualitative support with grid-free visualizations of material, geometry, and process dependencies, aiding in design decisions. The presentation concluded with an outlook on extending PINNs for more complex problems, including CAD parts, meshed components, inhomogeneous and anisotropic materials, and process-induced material variations. This research paves the way for more comprehensive simulations in composite manufacturing and other engineering fields.