Researchers from Tsinghua University and Fuzhou University have developed an innovative method for conducting full-range nonlinear analyses of structural systems using heterogeneous graph learning, as published in Engineering. The study introduces a computational model named StructGNN-N, which leverages deep learning techniques to significantly enhance the efficiency and accuracy of structural analysis, particularly in the context of digital twins.

Traditional methods for nonlinear structural analysis, primarily based on the finite-element (FE) framework, face challenges such as high modeling complexity and low computational efficiency. These limitations are particularly pronounced in modern engineering applications, including the digital twin context, which demands real-time updates and comprehensive mechanical information. In response, the research team proposed a novel approach integrating heterogeneous graph neural networks (GNNs) and sequence-to-sequence (Seq2Seq) models to address these challenges.

The core of this new method is the use of a heterogeneous graph (HetG) representation scheme, which can digitalize arbitrary structural systems with high fidelity. This scheme allows for the integration of various structural components and their connectivity into a unified graph structure. The researchers developed a composite feature learning framework comprising two main components: an HGNN-based module that encodes static features into embeddings with full structural semantics, and a Seq2Seq module that predicts history-dependent responses using these embeddings and external stimuli.

The computational model, StructGNN-N, was implemented and validated through numerical experiments involving real-world concrete structures. The results demonstrated that StructGNN-N successfully reproduces the full-range nonlinear responses of all nodes in the structure, with a computational efficiency level 1000 times greater than that of traditional elastoplastic history analysis using the FE method. This efficiency gain is crucial for practical engineering applications, especially in scenarios requiring rapid and accurate structural assessments.

Moreover, the study introduced a masked response training strategy to augment data utilization, addressing the issue of data scarcity in structural system analysis. By masking part of the target structural system’s information, this strategy enables significant data augmentation, enhancing the model’s ability to generalize across different structures and loading conditions.

The research also highlights the potential of the proposed method in the context of digital twins. The ability to conduct efficient and accurate nonlinear analyses makes StructGNN-N a promising tool for integrating into digital twin paradigms, where real-time updates and comprehensive mechanical information are essential for safety assessments and decision-making.

The study presents a significant advancement in the field of civil engineering by offering a more efficient and accurate method for nonlinear structural analysis. The integration of deep learning techniques, particularly heterogeneous graph learning and sequence processing, paves the way for future developments in smart structural engineering and digital twin applications.

The paper “Efficient Full-Range Nonlinear Analyses of Structural Systems Based on Heterogeneous Graph Learning,” is authored by Linghan Song, Chen Wang, Jiansheng Fan. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.07.001. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.