HybridSolver

Computer simulations are an essential tool in material, component and process development. However, physics based simulations quickly reach their limits in terms of both computing time and model size. This is especially true for fiber-reinforced-plastics (FRP), where physical phenomena extend over several orders of magnitude - from the micrometer scale at the fiber level to the component level in the meter range. The “HybridSolver” project hybridizes conventional numerical methods with physics-informed neural networks (PINNs) in order to efficiently overcome this challenge of multiscale simulation.

The hybridization is based on the coupling of two spatial scales. At the microscale (figure 1), partial differential equations are approximated by PINNs, while numerical solvers are used at the mesoscale. Both approaches are linked by multiscale techniques from materials science and treated as a non-standard optimal control problem. The implementation is carried out using parallelization and decomposition strategies, which ensures an efficient solution that is also resilient to inaccurate training data. (figure 2)

The “HybridSolver” project is being carried out in an interdisciplinary consortium consisting of the Weierstraß-Institut für Angewandte Analysis und Stochastik, the Leibniz-Institut für Verbundwerkstoffe and the Chair of Scientific Computing at the University of Kaiserslautern-Landau. Additionally, BOSCH and COMSOL are involved as industrial advisory board members.

The aim of the project is the hybridization of numerical methods with advanced physics-informed neural networks (PINNs) for the efficient calculation of complex multiscale simulations in the field of fiber-plastic composites.