Enhancing digital twin performance through optimizing graph reduction of finite element models.

Abstract

This research compares Graph Neural Networks reduction techniques for computational modeling in structural engineering. The proposed Finite Element Shape Logic Graph technology serves as a benchmark, capturing intricate details through fully connected logic graphs derived from Finite Element Models. Multiple graph reduction methods are introduced, each addressing computational efficiency: shortest path connectivity, Laplacian Matrix for Dimensionality Reduction, and an approach exploiting graph sparsity. This research contributes to structural engineering and offers insight for diverse applications in the evolving landscape of digital twins, particularly within the specific context of graph optimization and the efficiency of physical neural networks. Our proposed techniques are tested on datasets consisting of four different finite element models. Our study suggests that, while the shortest path method is a promising starting point for graph reduction in mechanical system modeling, leveraging Laplacian spectral reduction techniques is essential for achieving optimal solutions, particularly for larger finite element models. This approach served as the basis for the development of digital twins of mechanical structures, ensuring accurate representation and efficient analysis of complex engineering systems.