Structural partitioning for parallel construction of geometric nonlinear reduced-order models

Abstract

High-fidelity finite element models are widely used to predict the mechanical behavior of structures with complex geometries. While these models provide accurate results, they often require significant computational time, particularly when predicting nonlinear dynamic behavior. To address this, model order reduction techniques have been developed to reduce the computational time. However, constructing non-intrusive reduced-order models from high-fidelity finite element models still requires considerable computational time. This paper proposes a fully parallel process to accelerate the construction of geometrically nonlinear reduced-order models. In this approach, the structure is divided into multiple partitions, each assigned to a separate processor. The reduction basis for each partition ensures displacement consistency at partition interfaces without requiring additional modal coordinates at these interfaces. The parallel process operates without inter-processor communication, making it robust and straightforward to implement. It is compatible with various non-intrusive model order reduction techniques and achieves high computational efficiency. Notably, this approach introduces no additional errors, i.e., the reduced-order model constructed through the parallel process is identical to that obtained via traditional methods.