Developing a Silicon Beam Emulator Using Graph Attention Networks

Abstract

This study pioneers the application of Graph Attention Networks (GAT) and Graph Neural Networks (GNN), to the emulation of MEMS silicon beams under external loadings, representing a significant advancement in MEMS simulation. The novel augmented graph generation technique employed in this research enhances model performance and computational efficiency compared to traditional GNN message passing, particularly in data augmentation for training robust GAT. Silicon beam simulations were conducted under varying pressure conditions, and the results were found to be consistent with the theoretical solution, which can be a complementary roles of structural analysis combined with traditional finite element analysis (FEA). The trained GAT achieved a remarkable accuracy, with over 91% of predictions exhibiting an error no more than 3%, while the computation time for a single graph remained under 2 milliseconds. In the testing phase, the GAT-based computational simulation demonstrated a significant speed advantage, over 20 times faster than conventional FEA. The successful application of GAT in this context promises to reshape multiple aspects of technology development and optimization, paving the way for more efficient and accurate simulations across a wide range of disciplines.

Keywords

• Graph neural networks
• graph attention networks
• MEMS silicon beams
• finite element analysis
• augmented graph generation