A mesh-based geometric deep learning framework for rapid response prediction of large-scale and multi-component mechanical structures in engineering

Mesh-based finite element method (FEM) plays a critical role in simulating structural response. However, the analysis of complex physical processes, such as vehicle crashworthiness, is hindered by inherent high nonlinearities and large-scale meshes, leading to significant computational overhead and impeding rapid structure design. In recent years, the use of machine learning (ML) or deep learning (DL) methods to build surrogate models for simulations has gained much attention, which offers the potential to drastically reduce computational time while preserving accuracy. In this paper, we develop an end-to-end and mesh-based geometric DL framework that takes finite element (FE) solver files (such as.k file for LS-Dyna) containing mesh and material information as input, and quickly outputs response prediction of large-scale and multi-component mechanical structures in engineering, thus serving a promising alternative to FE solvers. We innovatively introduce the graph self-supervised learning (SSL) to transform FE data of structural component with varied material properties, complex geometric shapes and arbitrary number of unstructured meshes into low-dimensional embeddings, which are then employed to build an equivalent small-scale graph representation of the large-scale assembly, effectively alleviating the computational costs of subsequent prediction models. Then, we present GNN-FNN and GNN-Transformer models specifically designed for three different prediction tasks, including forecasting static and dynamic structural performance metrics, and constructing time-dependent physical fields. Using a large-scale industrial case of the electric vehicle (EV) under side pole impact, three regression tasks are carried out to assess the effectiveness of the proposed approach. Results reveal that the non-parametric model, free from the need for manually defined explicit parameters, excels in extracting implicit parameters for diverse structures, which support satisfactory prediction accuracy in each task with a considerable speedup than the simulation. Besides, it is surprising that our model is weakly sensitive to the moderate variation in the mesh resolution, which is valuable for practical engineering applications. The adaptability and scalability of our method are further verified on three additional industrial cases with varied structural simulation scenarios and progressively increasing FE model complexity. This work offers an effective surrogate model to accelerate the response evaluation of mechanical structures in engineering and shorten the design cycle requiring iterative optimization.