Non-intrusive coupling with local deep learning models for the effective simulation of welded structures in explicit dynamics

Despite advancements in high-performance computing and new developments in numerical methods, solving large structural problems featuring multiple complex localized behaviors remains a significant challenge. For this reason, in the automotive industry, highly simplified Finite Element Method (FEM) models are used for crash simulations, using 1D elements to represent spot welds. Modeling with thousands of spot welds with refined elements not only increases the degrees of freedom but also reduces the time step used in explicit schemes. To address this challenge, both intrusive and non-intrusive Domain Decomposition Methods (DDMs) have been developed in the past, where the localized features are solved separately on their own time scales. This thesis aims to develop a non-intrusive coupling strategy that integrates a data-driven Reduced Order Model (ROM) of a local model into an explicit dynamic global solver. The research comprises three phases. Initially, a Physics-Guided Architecture of Neural Networks (PGANN)-based reduced model for the local problem is developed. This model is then integrated into the global model through a non-intrusive local/global coupling method. Lastly, an error estimation technique is formulated to identify potential localized regions where NN-based enhancements are required. The proposed method is demonstrated on various examples with increasing complexity, ranging from a simple 2D plate problem to 3D industrial structures with multiple spot welds.