48: Evaluating Zoning Granularity In Graph Convolutional Networks For Predicting Energy and Structural Performance

Early-stage building design often treats energy, structure, and space as disconnected domains, limiting the potential for holistic, low-carbon solutions. To address this, we explore graph-based machine learning for considering these domains simultaneously while working at the plan level. We propose predicting structural material quantity (SMQ) and energy use intensity (EUI) using spatially grounded Graph Convolutional Networks (GCNs). The framework is demonstrated on a parametric case study of a floorplan with a central atrium. Three levels of spatial granularity were tested, ranging from detailed zone/component graphs to highly abstracted configurations. High-resolution graphs achieved strong predictive accuracy (R² > 0.97), while simplified graphs underperformed due to limited relational information. Compared to baseline regression models, which captured energy trends but failed to predict structural performance, GCNs effectively modeled spatial and load-dependent interactions. These findings examine graph granularity in early performance prediction and suggest future work on zone-level feature fusion to support multi-objective generative design.