Graph Neural Networks for Structural Displacement Prediction
Hung-Fu Chang, Tzu-Kang Lin, Yung-Li Cheng · 2026
Read Graph Neural Networks for Structural Displacement Prediction by Hung-Fu Chang, Tzu-Kang Lin, Yung-Li Cheng (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-08303.txt.
Introduction
Hung-Fu Chang, Tzu-Kang Lin, and Yung-Li Cheng predict how structures deform under load. They use a graph neural network instead of a full finite element run. Displacement prediction supports structural health monitoring and seismic safety checks. Finite element analysis is too slow for real-time use, though.
The team draws a frame as a graph. Joints are nodes, and members are edges. They feed in geometric and mechanical properties. The network then learns the load-to-response map from simulated data. Training data came from a two-story frame modeled in ANSYS. A conventional neural network served as the baseline. The graph model uses the structure's own links to make fast estimates.
Catalog record
| CREATOR | Hung-Fu Chang, Tzu-Kang Lin, Yung-Li Cheng |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.08303
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Graph Neural Networks for Structural Displacement Prediction Hung-Fu Chang, Tzu-Kang Lin, Yung-Li Cheng Abstract Accurate prediction of structural displacements under external loading is fundamental to structural health monitoring and seismic safety assessment. Although the finite element method (FEM) remains the prevailing approach because of its high accuracy, its considerable computational cost restricts its suitability for real-time monitoring applications. To address this limitation, this study proposes a data-driven framework based on Graph Neural Networks (GNNs), in which structural systems are represented as graphs with joints modeled as nodes and structural members as edges. By incorporating both geometric and mechanical properties into the graph representation, the proposed model learns the relationship between applied loads and structural responses directly from simulated data. A synthetic dataset was generated from a two-story frame structure using ANSYS, and both a conventional Neural Network (NN) and a GNN were trained for comparison. The results show that the proposed GNN framework predicts displacements and rotations with high accuracy and outperforms the NN model, demonstrating its potential as a fast and efficient alternative to traditional FEM-based analysis. K eywords GNN, Structural Displacement, Machine Learning Accurate prediction of structural displacements is a fundamental problem in structural engineering because displacement response is directly related to serviceability, damage evaluation, and the overall safety of structural systems subjected to external loads. In practice, reliable displacement estimates are essential for performance assessment under dynamic and quasi-static loading conditions, such as earthquakes, wind, and other operational demands. They are also critical for structural health monitoring, condition assessment, and the development of rapid decision-support systems for infrastructure management. For decades, structural response prediction has been dominated by analytical formulations and numerical simulation techniques, most notably the finite element method (FEM) [ 4 ] . FEM-based analysis follows a bottom-up strategy in which a structure or an object is discretized into a large amount of elements, local stress and strain are computed at the element level, and the global displacement response is subsequently assembled from these local interactions. Because of FEM method’s physical foundation and ability to model complex geometries, boundary conditions, and various material behaviors, It has become the standard tool for structural analysis and simulation. While FEM is robust and physically interpretable, high accuracy simulations typically demand substantial preprocessing effort, repeated numerical solution procedures, and considerable computational resources [ 11 ] . This challenge becomes even more pronounced for structures with intricate geometries or multiple constituent materials, where the complexity of the system significantly increases the cost of modeling and analysis. Consequently, their application becomes limited in scenarios that require rapid evaluation, repeated analysis over many loading cases, or near-real-time response prediction. Recent advances in machine learning have introduced a new paradigm for structural analysis, shifting the emphasis from purely physics-based computation to data-driven modeling [ 9 , 10 , 13 ] . Unlike conventional bottom-up approaches, machine learning methods consider the structure as an integrated system and learn the relationship between structural inputs, such as external loads, and structural responses, such as displacements, directly from data. In this framework, the structure is treated as an input–output system, without explicitly resolving the detailed behavior of individual materials at every stage of analysis. Once trained, the model can predict structural displacements under new loading conditions without repeatedly solving the governing equations, thereby significantly reducing computational cost. Moreover, such approaches offer strong potential for practical implementation in real-world applications, particularly when sensor systems are available to provide measurements of loads, displacements, and deformations. These advantages make machine learning especially attractive for applications requiring rapid prediction, scalability, and efficient repeated inference. Graph neural networks (GNNs), initially introduced in 2005 [ 2 ] , are designed to process data that can naturally be represented as graphs. They address the limitations of traditional machine learning methods in handling topological relationships and data structures [ 15 , 14 ] . Therefore, GNNs are particularly effective for problems in which the interactions among connected entities can be expressed in a graph. This characteristic makes GNNs particularly attractive for structural analysis, since a structural system can be naturally represented as a graph, where joints are treated as nodes and beams or columns as edges. Such a representation preserves the physical connectivity of the structure and allows the model to capture the influence of member-to-member interactions on the global response. Motivated by this perspective, this study investigates the application of graph neural networks (GNNs) to structural displacement prediction, a topic that remains relatively underexplored in existing literature. The main objective is to assess whether a GNN can achieve accurate and computationally efficient prediction of structural response in comparison with other machine learning models. To facilitate model training and evaluation, a synthetic data generation strategy is employed to produce training and evaluation datasets. Specifically, the nodal displacements and rotations of a two-story structural frame subjected to external loading are investigated within machine learning models. In this way, the study provides insights of using machine learning methods for structural analysis as an initial contribution to the growing body of research interest on machine learning-based structural analysis. We also contribute a dataset that can be used in the future machine learning based studies. Both point to the potential of such approaches for future applications in structural health monitoring and simulation. 1 Literature Review Existing applications of Graph Neural Networks (GNNs) in civil and structural engineering can generally be grouped into two main directions. The first focuses on mesh-based simulation and surrogate modeling for structural analysis. A representative study in this area is MeshGraphNets proposed by [ 10 ] , a general-purpose framework for mesh-based simulation that employs message-passing GNNs on irregular meshes to model physical systems efficiently. Their work addresses a solution for the high computational cost due to the high-dimensional numerical simulations. Align with this foundation, [ 13 ] proposed Physics-Informed MeshGraphNets (PI-MGNs), which integrate MeshGraphNet with physics-informed training through FEM/PDE constraints for non-stationary and nonlinear simulations on arbitrary meshes, further improving computational efficiency while preserving physical consistency. Similarly, [ 11 ] developed a GNN-based model as an alternative to FEM in order to reduce the time required for simulation. More recently, [ 5 ] extended MeshGraphNet by coupling it with a physics-attention Transformer to better capture long-range interactions, thereby addressing the limited efficiency of iterative message passing on large, high-resolution meshes. This improvement is particularly relevant for large-scale civil structures, where distant structural coupling may play an important role. The second major application of GNNs in this field concerns structural damage detection and health monitoring. [ 12 ] introduced a dual-graph framework in which one graph represents sensor topology and another capture feature correlations, combined with a CNN-based denoising and compression module for vibration signals. Their method demonstrated strong accuracy and robustness on benchmark datasets while addressing practical constraints in structural health monitoring (SHM). [ 1 ] proposed a graph representation that incorporates both structural geometry and measured vibration data, and applied Graph Convolutional Networks to exploit spatial correlations among sensor locations for damage identification. [ 6 ] developed a dynamic GNN (DynGNN) in 2024 for near-real-time damage identification, improving upon conventional approaches based on static graph structures. Their validation on test data from a steel truss bridge under vehicle loading suggested promising applicability to