Code-compliant optimal design of reinforced concrete slab beam systems for low to mid-rise buildings using heterogeneous graph neural networks with attention layers

Conventionally, the design of reinforced concrete (RC) shear frames is carried out by considering only a limited number of design alternatives, typically guided by rules of thumb and prescriptive code provisions. While such approaches are straightforward, they often result in material-intensive and suboptimal solutions. A more rational design process requires optimised beam-slab systems, which form a critical stage in RC shear frame design. This process must simultaneously satisfy structural safety, serviceability, and cost efficiency. However, it is inherently a high-dimensional and highly constrained optimisation problem that demands both advanced expertise in structural optimisation and significant computational resources. In this study, we present an optimal beam-slab system design workflow for buildings with rectangular and its derivative plan forms that combines a heterogeneous graph neural network and feed-forward neural networks, which not only generate optimal designs but also ensure code compliance. In this workflow, (a) a heterogeneous graph neural network is used to predict critical actions at structural nodes, beams, and slab elements under different loading scenarios; (b) two feed-forward neural networks predict optimal beam and slab reinforcement details for a given layout; and (c) for a given design domain, the combination of the graph neural network, feed-forward networks, and particle swarm optimisation predicts the optimal beam-slab layout (considering primary and secondary beams) along with optimal reinforcement detailing for beam-slab elements. Results show high accuracy, with the heterogeneous graph neural network achieving less than 4% Mean Absolute Percentage Error (MAPE) in action predictions, while the feed-forward networks achieved 3.5% MAPE for beam details and 6.1% for slab reinforcement. The workflow demonstrates a 30-fold increase in speed for optimal beam-slab system design compared to traditional optimisation methods. Case studies for various design examples highlight the method’s capability, achieving a 10-45% reduction in material costs compared to common rule-of-thumb designs. By enhancing computational efficiency and implementing a comprehensive optimisation scheme, this work facilitates sustainable RC design and minimises resource consumption. • A GNN based workflow for optimal reinforced concrete beam slab design • Predicts critical actions, optimal element details, and optimal beam slab layout • Delivers optimal, code-compliant beam and slab details with high accuracy • Generates the optimal details 30 times faster than conventional methods • Reduces material costs by 10-45% compared to conventional designs