Machine learning (ML) is increasingly used in structural engineering and design, yet its broader adoption is hampered by the lack of openly accessible datasets of structural systems. We introduce BridgeNet, a publicly available graph-based dataset of 20,000 form-found bridge structures aimed at enabling Graph ML and multi-modal learning in the context of conceptual structural design. Each datapoint consists of (i) a pinjointed equilibrium wireframe model generated with the Combinatorial Equilibrium Modeling (CEM) form-finding method, (ii) a volumetric 3D mesh obtained through forceinformed materialization, and (iii) rendered images from two canonical camera angles. The resulting dataset is modality-rich and application-agnostic, supporting tasks such as CEMspecific edge classification and parameter inference, surrogate modeling of form-finding, cross-modal reconstruction between graphs, meshes and images, and generative structural design. BridgeNet addresses a key bottleneck in data-driven applications for structural engineering and design by providing a dataset that facilitates the development of new MLbased approaches for equilibrium bridge structures.
💡 Deep Analysis
Deep Dive into Arxiv 2512.14496.
Machine learning (ML) is increasingly used in structural engineering and design, yet its broader adoption is hampered by the lack of openly accessible datasets of structural systems. We introduce BridgeNet, a publicly available graph-based dataset of 20,000 form-found bridge structures aimed at enabling Graph ML and multi-modal learning in the context of conceptual structural design. Each datapoint consists of (i) a pinjointed equilibrium wireframe model generated with the Combinatorial Equilibrium Modeling (CEM) form-finding method, (ii) a volumetric 3D mesh obtained through forceinformed materialization, and (iii) rendered images from two canonical camera angles. The resulting dataset is modality-rich and application-agnostic, supporting tasks such as CEMspecific edge classification and parameter inference, surrogate modeling of form-finding, cross-modal reconstruction between graphs, meshes and images, and generative structural design. BridgeNet addresses a key bottleneck in data-dr
📄 Full Content
Machine Learning (ML) is increasingly being adopted in the Architecture, Engineering and Construction (AEC) industry, including structural engineering and structural design. Among the various data representations used, graphs stand out as a particularly flexible and powerful medium for ML applications, naturally encoding discrete wireframe structures such as pin-jointed frameworks, discrete form-found structures, and structural frames. Correspondingly, Graph Machine Learning (Graph ML) methods, such as Graph Neural Networks (GNN), can operate directly on graph-structured data and have already been applied to a wide range of problems in structural engineering and design.
Bridge structures stand out as a particularly relevant use case for Graph ML applications: From a structural design perspective, bridges represent a suitable case study for cross-typological design due to their wide variety in topology and geometry of viable designs. From a structural engineering perspective, the substantial forces involved in bridge systems make them compelling candidates for advanced structural analysis and optimization. Moreover, the worldwide aging of existing bridge infrastructure underscores the urgency of developing efficient, data-driven methods for their assessment, rehabilitation, and new design.
Despite the widely recognized potential of ML applications in structural engineering and design, their adoption is not yet widespread. One of the factors inhibiting the further development of ML in these fields is the lack of generally available training data. At the same time, the most widely applied ML archetype, i.e., supervised learning, is critically dependent on high-quality training data to be able to make accurate predictions. Architecture and engineering firms are understandably reluctant to publicly share the data they have from their design portfolios. On the other hand, the data that is developed and used in academic research is also not always made publicly available. In general, preparing data so that it is accessible to others requires an additional investment of time and resources that may not always be available or prioritized. Instead, most ML models in structural engineering and design are trained on datasets that were specifically developed for the application of interest. This leads to duplicated efforts, hinders reproducibility, and complicates the making of meaningful comparisons across methods. The field, therefore, lacks broadly recognized, openly accessible datasets of wireframe structures that are neutral with respect to specific downstream applications, and rich enough to support a diverse range of data-driven tasks.
This paper introduces BridgeNet, a publicly available graph-based dataset of bridge structures intended for ML applications in structural engineering and design. Our focus is on pin-jointed bar structures that satisfy equilibrium through axial forces only, corresponding to the typical level of abstraction for the conceptual structural design stage. Structure are generated using a parametric model based on the Combinatorial Equilibrium Modeling (CEM) form-finding method (Ohlbrock and D’Acunto, 2020). We furthermore derive 3D mesh and 2D image representations from the wireframe graph-based structures, such that each datapoint in BridgeNet is represented in three different modalities. In this way, BridgeNet is positioned to facilitate a wide range of multi-modal ML applications for the conceptual design and analysis of bridge structures.
This paper is based on two main contributions:
A graph-based dataset of bridge structures (BridgeNet) of 20,000 bridges. For each bridge, the dataset includes: (i) a wireframe model of a pin-jointed structure in equilibrium, (ii) a 3D mesh model based on internal force magnitudes, and (iii) a 2D rendered image of the 3D model from two canonical camera angles. BridgeNet is hosted through Hugging Face and can be accessed at the following link: https://huggingface.co/datasets/lazlo-bleker/bridge-net
.
A data generation workflow based on form-finding, materialization, and rendering with which BridgeNet is generated, that also serves as a template for the generation of data of additional structural typologies.
While the availability of data for ML applications in structural design and engineering is generally limited, a few publicly available datasets of structures do exist. Particularly in the field of topology optimization, several image-based datasets are available (Sosnovik and Oseledets, 2019;Mazé and Ahmed, 2023;Bastos, 2025;Li et al., 2025). These datasets are typically generated by varying boundary conditions, such as loads and supports, as well as parameters of the optimization process, including the targeted volume fraction. While most of these datasets include 2D structures, Dittmer et al. published a dataset of 3D voxelized topology optimization results (Dittmer et al., 2023).
Topology optimization lends itself well to the creation