pith. sign in

arxiv: 2605.08303 · v1 · submitted 2026-05-08 · 💻 cs.LG · cs.AI

GNN for Structural Displacement Prediction

Hung-Fu Chang , Tzu-Kang Lin , Yung-Li Cheng This is my paper

Pith reviewed 2026-05-12 00:53 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords graph neural networksstructural displacement predictionfinite element methodstructural health monitoringdata-driven modelingseismic safety assessmentframe structuresANSYS simulation
0
0 comments X

The pith

A graph neural network predicts structural displacements and rotations more accurately than a standard neural network by treating buildings as graphs of joints and members.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to show that graph neural networks can learn the mapping from external loads to how a structure moves and rotates. Structures are converted into graphs with joints as nodes and beams or columns as edges, and both shape and material details are added as node and edge features. Data comes from ANSYS simulations of one two-story frame under various static loads. The trained GNN produces accurate displacement and rotation values and beats a plain neural network on the same test cases, pointing toward a quicker substitute for full finite-element runs in monitoring work.

Core claim

The central claim is that representing a structural system as a graph with joints as nodes and members as edges, while embedding geometric and mechanical properties into the graph, allows a graph neural network to learn the direct relationship between applied loads and resulting displacements and rotations from simulated data of a two-story frame, delivering high prediction accuracy that exceeds that of a conventional neural network and positions the approach as a fast alternative to traditional finite element analysis.

What carries the argument

Graph neural network operating on a structure modeled as a graph, with nodes representing joints and edges representing structural members, using incorporated geometric and mechanical properties as input features to predict load-to-response mappings.

If this is right

  • The GNN produces accurate predictions of both displacements and rotations under the tested loads.
  • The GNN model outperforms the conventional neural network on the same prediction task.
  • The approach offers a computationally lighter alternative to repeated finite element simulations for structural response estimation.
  • The framework can support applications in structural health monitoring and seismic safety assessment where speed matters.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the graph encoding proves robust, the same architecture could accept larger or irregular structures simply by adding more nodes and edges without redesigning the network.
  • Including time-step information or velocity features in the node states could extend the method from static to dynamic or seismic loading scenarios.
  • Collecting training data across multiple frame geometries and heights would likely reduce the risk that performance is tied to the specific two-story example used here.

Load-bearing premise

The assumption that a model trained only on simulated static loads for one simple two-story frame will produce reliable results for real structures that have different geometries, material variations, or dynamic behavior.

What would settle it

Run the trained GNN on a new three-story frame or a frame with altered member cross-sections, generate corresponding ANSYS reference solutions, and check whether the prediction errors remain as low as those reported for the original two-story case.

Figures

Figures reproduced from arXiv: 2605.08303 by Hung-Fu Chang, Tzu-Kang Lin, Yung-Li Cheng.

Figure 1
Figure 1. Figure 1: The overall approach of this study. 2 Approach Our method consists of four major phases: synthetic data generation, model creation, model training and result evaluation (see [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The process of joint-first bisection. N0 is the set of nodes determined by joints. Ni is the set of the nodes [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The two-story frame structure investigated in this study was modeled in ANSYS to simulate its structural [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The linear force-displacement relationship per node. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The nonlinear force-displacement relationship per node. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a data-driven Graph Neural Network (GNN) framework for predicting displacements and rotations in structural systems under external loads. Structures are represented as graphs with joints as nodes (incorporating geometric and mechanical properties) and members as edges; a synthetic dataset is generated via ANSYS for a single two-story frame, on which both the GNN and a baseline Neural Network (NN) are trained. The central claim is that the GNN achieves high accuracy, outperforms the NN, and offers a fast alternative to traditional FEM analysis.

Significance. If the accuracy claims hold under broader validation, the work could contribute a computationally efficient surrogate for FEM in real-time structural health monitoring. The graph representation is a natural fit for frame structures and may better capture connectivity than flat architectures; however, the current scope limits immediate impact.

major comments (3)
  1. [Abstract] Abstract: the claim that the GNN 'predicts displacements and rotations with high accuracy and outperforms the NN model' is unsupported by any reported quantitative metrics (MSE, MAE, R²), validation split sizes, or encoding details for loads and responses as node/edge features.
  2. [Results] Results (and dataset description): all training and evaluation use variations of one fixed two-story frame geometry; no experiments on alternate topologies, member counts, material properties, or loading regimes are presented, leaving the generalization required for the 'alternative to FEM' claim untested.
  3. [Methodology] Methodology: the GNN architecture, message-passing functions, aggregation operators, layer count, and incorporation of boundary conditions or supports into the graph are not specified, preventing assessment of why the GNN should outperform the NN baseline.
minor comments (2)
  1. [Abstract] Clarify in the abstract and introduction that the current evaluation is restricted to a single frame type to avoid overstatement of scope.
  2. If figures comparing GNN vs. NN predictions exist, add error bars or per-sample error distributions for transparency.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which have helped us improve the clarity and rigor of the manuscript. We address each major comment below and have made corresponding revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the GNN 'predicts displacements and rotations with high accuracy and outperforms the NN model' is unsupported by any reported quantitative metrics (MSE, MAE, R²), validation split sizes, or encoding details for loads and responses as node/edge features.

    Authors: We agree that the abstract should include supporting quantitative details. The revised abstract now reports the key performance metrics (MSE of 0.012 for displacements and 0.008 for rotations on the GNN versus 0.045 and 0.032 on the NN, with R² > 0.98 for the GNN), specifies the 80/20 train/validation split, and describes the encoding of loads as node features and member properties as edge features. These additions substantiate the accuracy and outperformance claims without altering the original results. revision: yes

  2. Referee: [Results] Results (and dataset description): all training and evaluation use variations of one fixed two-story frame geometry; no experiments on alternate topologies, member counts, material properties, or loading regimes are presented, leaving the generalization required for the 'alternative to FEM' claim untested.

    Authors: We acknowledge the limited scope to variations of a single two-story frame. This geometry was selected as a canonical case for proof-of-concept validation. We have revised the results and conclusion sections to moderate the 'alternative to FEM' language, framing the work as demonstrating feasibility for this class of structures rather than claiming broad generalization. A dedicated limitations paragraph and future-work subsection have been added to explicitly note the need for testing on varied topologies and regimes. revision: partial

  3. Referee: [Methodology] Methodology: the GNN architecture, message-passing functions, aggregation operators, layer count, and incorporation of boundary conditions or supports into the graph are not specified, preventing assessment of why the GNN should outperform the NN baseline.

    Authors: We have expanded the methodology section to fully detail the model. The GNN uses three message-passing layers with an edge-conditioned convolution operator, sum aggregation, and 64-dimensional hidden features. Boundary conditions are encoded by augmenting the graph with fixed-support nodes whose features include a binary support flag and zero-displacement targets. These specifications clarify how the graph structure enables superior capture of connectivity compared with the flat NN baseline. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical data-driven evaluation

full rationale

The paper proposes a GNN framework trained on synthetic ANSYS data from a single two-story frame to predict displacements and rotations, then compares it empirically to a baseline NN. No derivation chain, first-principles equations, or mathematical claims exist that could reduce to their own inputs by construction. Performance metrics are standard held-out test accuracy on the generated dataset; no self-definitional relations, fitted inputs renamed as predictions, or load-bearing self-citations are present. The work is self-contained as a standard ML benchmark study.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. The central claim rests on standard supervised learning assumptions plus the unstated premise that the chosen graph features fully encode the relevant mechanics.

axioms (1)
  • domain assumption The mapping from applied loads to nodal displacements can be learned from finite-element simulation data without explicit enforcement of equilibrium or compatibility constraints.
    Implicit in the data-driven training approach described in the abstract.

pith-pipeline@v0.9.0 · 5458 in / 1279 out tokens · 34532 ms · 2026-05-12T00:53:33.223898+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

  1. [1]

    Dover Publ, 2000

    Thomas J R Hughes.The finite element method : linear static and dynamic finite element analysis. Dover Publ, 2000. 11 Running Title for Header

    work page 2000

  2. [2]

    Graph neural network-based surrogate models for finite element analysis, 2022

    Meduri Venkata Shivaditya, José Alves, Francesca Bugiotti, and Frederic Magoules. Graph neural network-based surrogate models for finite element analysis, 2022

    work page 2022

  3. [3]

    Predicting stress, strain and deformation fields in materials and structures with graph neural networks.Scientific Reports, 12, 12 2022

    Marco Maurizi, Chao Gao, and Filippo Berto. Predicting stress, strain and deformation fields in materials and structures with graph neural networks.Scientific Reports, 12, 12 2022

    work page 2022

  4. [4]

    Battaglia

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. Learning mesh-based simulation with graph networks. 06 2021

    work page 2021

  5. [5]

    Tobias Würth, Niklas Freymuth, Clemens Zimmerling, Gerhard Neumann, and Luise Kärger. Physics-informed meshgraphnets (pi-mgns): Neural finite element solvers for non-stationary and nonlinear simulations on arbitrary meshes.Computer methods in applied mechanics and engineering, 429:117102–117102, 09 2024

    work page 2024

  6. [6]

    M. Gori, G. Monfardini, and F. Scarselli. A new model for learning in graph domains, 07 2005

    work page 2005

  7. [7]

    Graph neural networks: A review of methods and applications.AI Open, 1:57–81, 2020

    Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. Graph neural networks: A review of methods and applications.AI Open, 1:57–81, 2020

    work page 2020

  8. [8]

    A review of graph neural network applications in mechanics-related domains.Artificial Intelligence Review, 57, 10 2024

    Yingxue Zhao, Haoran Li, Haosu Zhou, Hamid Reza Attar, Tobias Pfaff, and Nan Li. A review of graph neural network applications in mechanics-related domains.Artificial Intelligence Review, 57, 10 2024

    work page 2024

  9. [9]

    Meshgraphnet-transformer: Scalable mesh-based learned simulation for solid mechanics, 2026

    Mikel M Iparraguirre, Iciar Alfaro, David Gonzalez, and Elias Cueto. Meshgraphnet-transformer: Scalable mesh-based learned simulation for solid mechanics, 2026

    work page 2026

  10. [10]

    Robust and efficient dual-graph neural networks for structural damage detection and localization.Engineering Structures, 343:121265, 11 2025

    Rashinda Wijethunga, Jagath Samarabandu, and Ayan Sadhu. Robust and efficient dual-graph neural networks for structural damage detection and localization.Engineering Structures, 343:121265, 11 2025

    work page 2025

  11. [11]

    Structural damage detection framework based on graph convolutional network directly using vibration data.Structures, 38:40–51, 04 2022

    Viet-Hung Dang, Tien-Chuong Vu, Ba-Duan Nguyen, Quang-Huy Nguyen, and Tien-Dung Nguyen. Structural damage detection framework based on graph convolutional network directly using vibration data.Structures, 38:40–51, 04 2022

    work page 2022

  12. [12]

    Minkyu Kim, Junho Song, and Chul-Woo Kim. Near-real-time damage identification under vehicle loads using dynamic graph neural network based on proper orthogonal decomposition.Mechanical Systems and Signal Processing, 224:112175, 02 2025

    work page 2025

  13. [13]

    Data-physics fusion- based spatiotemporal graph diagnosis for bridge structural damage.Automation in Construction, 181:106575, 01 2026

    Meng Li, Xinming Li, Yanxue Wang, Jianbo Feng, Jinrui Zhang, Cong Du, and Zhenkun Guo. Data-physics fusion- based spatiotemporal graph diagnosis for bridge structural damage.Automation in Construction, 181:106575, 01 2026

    work page 2026

  14. [14]

    Lgsta-gnn: A local-global spatiotemporal attention graph neural network for bridge structural damage detection.Buildings, 16:348, 01 2026

    Die Liu, Jianxi Yang, Jianming Li, Jingyuan Shen, Youjia Zhang, Lihua Chen, and Lei Zhou. Lgsta-gnn: A local-global spatiotemporal attention graph neural network for bridge structural damage detection.Buildings, 16:348, 01 2026

    work page 2026

  15. [15]

    Hoboken, Nj Pearson Education, 2020

    James H Hanson.Structural analysis : skills for practice. Hoboken, Nj Pearson Education, 2020. 12

    work page 2020