SimTech researchers Jonas Kneifl and Jörg Fehr from the Institute of Engineering and Computational Mechanics (ITM) at the University of Stuttgart, along with Steven L. Brunton and J. Nathan Kutz from the University of Washington, have developed a new multi-hierarchical surrogate learning framework for efficient and accurate modeling of three-dimensional dynamical systems.
This novel approach, detailed in their paper titled "Multi-Hierarchical Surrogate Learning for Explicit Structural Dynamical Systems Using Graph Convolutional Neural Networks," uses graph convolutional neural networks (GCNNs) to build surrogate models at various resolution levels, enhancing the efficiency and accessibility of simulations.
Jonas Kneifl and Jörg Fehr collaborated with Brunton and Kutz to tackle the computational challenges posed by traditional high-fidelity finite element simulation models. While these models offer high detail, they require extensive computational resources, limiting their practical applications.
The new framework addresses this issue by employing a hierarchical approach that captures global dynamics on coarse models and refines local details on finer ones. This transfer learning technique not only speeds up the learning process but also allows the creation of adaptable models suited for different hardware and accuracy needs.
In their study, they demonstrated the application of their method using a racing kart model, which serves as a proxy for industrial crash simulations. They showed that their multi-hierarchical framework could accurately reproduce the complex dynamics of a frontal impact scenario, encompassing both large-scale and small-scale phenomena by transferring learned behavior from coarse to finer levels.
This method stands out by generating multiple surrogate models for the same system, each with varying hardware requirements and increasing accuracy. This approach not only reduces computational effort but also makes high-fidelity finite element simulations more accessible for real-time applications and environments with limited computing capacity.
The work of Kneifl, Fehr, Brunton, and Kutz represents a significant advancement in computer-aided engineering, enabling high-accuracy simulations with reduced computational effort.
The study is currently available as a preprint on arXiv: https://arxiv.org/abs/2402.09234.