Knowledge‐enhanced graph neural networks for construction material quantity estimation of reinforced concrete buildings

Abstract

The construction material quantity (CMQ) is widely concerned in the structural design of reinforced concrete buildings and is often included among the objective functions of computer‐aided optimization design techniques. To minimize construction cost and carbon emissions, an accurate and efficient CMQ estimation method is timely required. In this study, a novel graph neural network (GNN) is proposed, whose architecture and loss function are specifically designed for CMQ estimation. With a heterogeneous feature fusion mechanism, the GNN can automatically extract features from all CMQ‐related information, in contrast to the existing data‐driven methods that rely heavily on manually selected features. By further incorporating a prior knowledge inclusion strategy, the GNN can avoid fundamental errors that might be encountered by purely data‐driven methods. To enrich the diversity of the CMQ dataset, a data augmentation method is proposed incorporating generative adversarial networks and parametric modeling. Numerical experiments and case studies show that the proposed CMQ estimation method is superior to the existing data‐driven methods in terms of accuracy and is 500 times faster than typical commercial structural design software. This study is anticipated to benefit the objective evaluation of computer‐aided design, thereby facilitating the promotion of low‐cost and low‐carbon‐emission building designs.

Hybrid surrogate model combining physics and data for seismic drift estimation of shear‐wall structures

“…Details of the five modules can be found in the previous work of the authors. 32 By adopting this parametric modeling procedure, the natural periods and maximum IDRs under the design earthquakes were obtained for each sample. According to the Chinese code, shear-wall structures are required to satisfy IDR max ≤ 0.1% under design earthquakes.…”

“…To estimate the seismic responses of buildings, it is necessary to consider their structural layouts (topological information), component dimensions (geometric information), and design conditions (textual information). To fully integrate the assorted input data mentioned above, a recently proposed GNN named Graph-GEN 32 was employed. This approach incorporates feature representation, feature aggregation, graph representation, and feature fusion algorithms specifically designed for building structure-related tasks, as shown in Figure 7.…”

“…Compared with other data-driven methods such as the ANN and CNN, Graph-GEN has proven to be a superior surrogate model for building structures. 32…”

“…Through its feature fusion mechanism, Graph-GEN can establish a mapping relationship between input and output data in a data-driven manner. 32 Because the physical quantities to be predicted (i.e., natural period, stiffness, and IDR) are inherently positive values, the loss function and output layer of Graph-GEN were modified to ensure that its predictions were 100% positive, as shown in Equations ( 2)- (4).…”

“…To fulfill the design optimization requirements, the structural design and design conditions of the shear-wall structures were represented as a six-channel pixel image, that is, a 512 × 256 × 6 input feature matrix. 32 For each channel (i.e., a 512 × 256 matrix), non-zero matrix elements indicate the presence of structural components, whereas zero matrix elements indicate the opposite. For the six channels, the values of the nonzero matrix elements represent the component width, component height, number of stories (𝑛 s ), design seismic acceleration (𝑎 d ), site characteristic period (𝑇 g ), and scaling factor (used to scale the structural layout into a 512 × 256 matrix).…”