Rapid Structural Analysis Method of a Novel Semi-Rigid Formwork Support System Using Graph Neural Networks

Le résumé fourni par la source

Objective Rapid prediction of structural responses is important for construction-stage checking and parametric analysis of Novel Semi-Rigid Formwork Support Systems (NSRFSSs), whose span number, span length, construction load, and joint rotational stiffness may vary considerably. Repeated finite element analysis (FEA) is computationally expensive, while conventional deep learning models are not well suited to structures with variable topology and size. A graph-neural-network-based model, termed Semi-Rigid Support Graph Neural Network (SRSGNN), was therefore developed to efficiently predict nodal Mises stress and vertical deflection.Methods A parametric finite element model of the NSRFSS was established in Abaqus. Horizontal beams and vertical supports were modeled using beam elements, while early-stripping heads were represented by connector elements with prescribed rotational stiffness. Static analysis was used to obtain stress and displacement responses, and linear elastic eigenvalue buckling analysis was conducted for representative cases. A total of 2 400 finite element samples were generated by varying span numbers in two directions, span length, concentrated load, line load, and joint rotational stiffness, while the support height was fixed at 3.1 m. Nodal Mises stress and vertical displacement (U_3) were extracted as learning targets. Among the samples, 2 000 were used for training, validation, and interpolation testing, and 400 larger-scale samples were used for extrapolation testing. A fine-grained graph representation, termed VNodeAsNode, was proposed by inserting virtual nodes along structural members to represent local responses within members. Node features included coordinates, boundary conditions, loads, material properties, and sectional characteristics, while edge features included connectivity, geometric length, edge type, and rotational stiffness. Based on this representation, SRSGNN was constructed using a graph isomorphism network with edge features (GINE). Node and edge information was propagated through multiple message-passing layers to predict nodal stress and deflection. Mean squared error was used as the loss function, and average relative accuracy was adopted as the main evaluation metric.Results and DiscussionsSRSGNN reproduced the main stress and deflection distributions obtained from FEA. The average relative accuracies for Mises stress and deflection were 92.33% and 90.51%, respectively, and the predicted high-response regions were generally consistent with the finite element results. For the 400 larger-scale extrapolation samples, the average relative accuracy remained above 90%, indicating that the model could accommodate changes in structural scale, node number, member number, and topology. The VNodeAsNode representation improved the description of local responses by introducing virtual nodes within members, while rotational stiffness as an edge feature enabled semi-rigid joint behavior to participate directly in message passing. For a representative case, one Abaqus analysis required about 42.1 s, whereas one SRSGNN prediction required about 0.0021 s, corresponding to a computational speed increase of more than 2.0×104 times. This efficiency supports rapid comparison of large numbers of structural configurations and preliminary structural optimization. The present model was limited to static responses under vertical construction loads. The dataset was generated from finite element models, the support height was fixed at 3.1 m, and semi-rigid joints were represented by linear rotational stiffness. Experimental validation and the effects of variable support height, joint nonlinearity, initial imperfections, geometric nonlinearity, and material nonlinearity require further study.Conclusions A graph-neural-network-based framework was developed for rapid prediction of local responses of NSRFSSs. The VNodeAsNode representation enabled variable-scale structures to be described in a unified graph form, while geometric, topological, loading, and semi-rigid connection information was incorporated into node and edge features. SRSGNN achieved average relative accuracies of 92.33% for Mises stress and 90.51% for deflection, maintained an accuracy above 90% for larger-scale extrapolation cases, and reduced the prediction time from about 42.1 s to 0.0021 s for the representative case. The method provides an efficient surrogate for repeated finite element analyses in rapid structural checking, parametric comparison, and preliminary optimization.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.