Automatic clash avoidance in steel reinforcement design using explainable graph neural networks and rebar embedding learning

Abstract

Steel reinforcement design is essential for the structural integrity and durability of reinforced concrete (RC) structures. However, rebar clashes frequently occur due to conventional design processes lacking precise bar positioning, leading to time-consuming and error-prone onsite modifications. Existing 3D analysis tools for clash detection are unsuitable for rebar design, which must comply with structural analysis and regional specifications. Therefore, this paper proposes an automatic and proactive rebar clash avoidance approach using graph neural networks (GNN) and rebar embedding learning. Vector and graph representations are introduced to model clash scenarios, while a GNN-based diagnosis framework detects clashes and classifies them as solvable or unsolvable. For unsolvable clashes, explainable GNN identifies the underlying causes, while Rebar2Vec generates optimal design alternatives to improve feasibility. Solvable clashes are resolved using multi-objective optimization, ensuring compliance with building codes. Experimental results demonstrate the approach's effectiveness in generating clash-free rebar layouts at the design stage.