From geometry to graph: Automation of building performance modeling via convex graph encoding

Early-stage design decisions strongly influence a building’s lifetime energy performance, yet the transformation from architectural geometry to analysis-ready models remains fragmented. This paper presents CUGER, an automated framework that converts complex 3D building geometries into physically consistent heterogeneous graphs for building energy analysis. CUGER integrates a convex optimization-based segmentation algorithm with structured graph encoding to directly transform CAD geometries into simulation- and learning-ready representations. Evaluated on 266 architecturally diverse building models, the framework achieved an average geometric-to-graph conversion success rate above 80% with strong topological consistency ( R 2 = 0 . 82 ). Building on the generated graphs, a heterogeneous ZoneGNN model enables zone-level energy prediction of single building, achieving accurate and stable performance ( R 2 = 0 . 88 ± 0 . 01 ) and outperforming MLP baselines, particularly under limited training data. Overall, CUGER establishes a topology-preserving and fully automatable bridge between architectural modeling, performance simulation, and data-driven prediction for early-stage building design. • Heterogeneous building info encoded in a single graph for simulation and prediction. • Convex optimization-based segmentation ensures geometric and topological fidelity. • ZoneGNN predicts hourly energy, outperforming MLPs, especially with limited data. • Framework automates graph construction, preserves physics, and supports design.