.png){kind=icon}
.png){kind=icon}
Return
Graph-based data-physics hybrid surrogate modeling for seismic response estimation of steel moment frame building portfolios
.png){kind=icon}
PRE
AI
S
DOI：10.1007/s10518-026-02629-z.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Regional-scale seismic damage assessment has traditionally relied on simplified models that use general building descriptions, resulting in inaccurate outcomes due to oversimplified correlations and inadequate data representation. While the increasing availability of detailed building information from digital twins and urban surveys enables higher-fidelity modeling, the computational cost and speed of finite-element analysis (FEA) remain impractical for large-scale applications. On the other hand, machine learning surrogate models are increasingly used to predict the seismic performance of buildings. However, traditional data-driven surrogate models typically rely on simple fixed-length vector-based features, which are insufficient to accurately represent the complex topology and characteristics of building structures, and consequently affect the accuracy of the predicted structural behavior. This paper proposes a graph-based hybrid surrogate modeling framework for estimating the seismic response of steel moment-frame building portfolios. The framework comprises three surrogate modeling approaches with different levels of data-physics integration, including two approaches that enhance multi-degree-of-freedom (MDOF) model-based response simulation through graph-based learning of graph-structured building data, and a physics-aware end-to-end graph neural network (GNN) approach for seismic response estimation. Results demonstrate that the proposed approaches significantly improve the estimation of structural responses compared to traditional parameterization-based MDOF modeling, reducing errors by more than 50%, and also outperform commonly used models. These findings highlight the potential of graph-based data-physics hybrid surrogate modeling to enable accurate and computationally efficient regional-scale seismic assessment using increasingly available detailed structural information from urban-scale databases and digital twins.
Keywords:
Graph neural networks
Hybrid modeling
Seismic analysis
Surrogate model
Multi-degree-of-freedom model
Steel moment frames
Journal
IF:
4.1
Papers:
3.6K
Citations:
1.0W
Organization
Cited Papers