A data-driven approach using Neural Networks for real-time modeling and simulation of structures with many welded points under impact

Large structural problems with complex localized behaviour are extremely difficult to solve. In the past, intrusive and non-intrusive domain decomposition methods (DDMs) have been developed to address this challenge. To further reduce computational time, a data-driven neural network-based metamodeling of the local scale is proposed here, which could establish an efficient and accurate mapping from interface velocities to interface forces and predict their time evolution, in the context of dynamics problems. It is difficult to obtain direct input-output relationships in explicit dynamics because these quantities are highly noisy by nature. To address this fundamental problem, we develop a new architecture called Physics-Guided Neural Networks (PGNN) . The key idea is to inject high-fidelity simulation quantities (such as displacement, stress, strain, and so on) from the local domain between the input and output layers of NN to improve learning within a solution space. In this study, we only inject one of these quantities, displacement. It can also account for unmodeled physics with fewer experimental data by imposing fundamental solid mechanics principles. The proposed method is exemplified by a spotwelded plates undergoing rapid deformation. The neural networks are trained with explicit FEM solver-generated simulation data. The PGNN results are in good agreement with the FEM solutions for both the training dataset and the unseen dataset test cases. At the conference, a preliminary implementation of a local/global coupling framework for explicit dynamics with a developed metamodel will be presented.