Non-intrusive coupling of neural network-based local model and explicit dynamics

In solving large structural problems with multiple complex localized behaviors, current industry practices rely on highly simplified Finite Element Method (FEM) models, leading to inaccurate failure predictions. Various domain decomposition techniques have emerged to tackle these challenges, with non-intrusive coupling methods standing out in industrial applications due to their flexibility. Despite their advantages, these methods typically require numerous iterations between local and global problems, posing significant computational challenges. To address this issue, the article proposes substituting the FEM-based local problem with a Neural Network-based Reduced Order Model (ROM). The work focuses on non-intrusive coupling, integrating the ROM with an explicit dynamic solver, resulting in significant reduction in computational time. The effectiveness of the proposed method is shown through two numerical examples, each increasing in complexity, to showcase the methods adaptability and scalability.