Generative design for modular construction structures based on expert knowledge-informed graph neural networks and meta-heuristics

With the growing demands in the construction industry, coupled with cost pressures and environmental concerns, modular building (MB) solutions are proposed to address the challenges. However, the design process of MB is more fragmented and complex, especially the structural design. This requires a reconsideration of the automated approach for its layout design, incorporating structural design. This paper develops a generative AI-enabled framework, focusing on the structural design of reinforced concrete MB. The proposed hybrid approach integrates a graph neural network-based model in a genetic generative design framework to surrogate structure design. And multiple structural related objectives are optimized in this framework. It was tested in a real project in Hong Kong and compared with the engineer's design. The optimal Pareto-balanced compromise solution resulted in a 12 % increase in usable floor area, a 7 % increase in structural performance, and a 23 % reduction in construction cost.