Design Model Transformation Testing Datasets for Architecture Early Design

This repository provides benchmark datasets for validating the performance of the proposed model-transformation framework in early-stage architectural design. In early design practice, models typically include only primary structural elements and apertures, with implicit but well-defined spatial organization and minimal semantic annotation. These characteristics make early design models highly flexible, yet challenging for automated interpretation and downstream analysis.
The transformation evaluated using these datasets consists of two sequential steps. First, unstructured boundary-representation (B-rep) design models are transformed into ontology-based Knowledge Graph (KG) or Labeled Property Graph (LPG) representations, where building elements, spaces, and their topological relationships are explicitly extracted and encoded. Second, the ontology models are further transformed into Building Energy Models (BEMs), enabling the derivation of performance-related metrics. Together, these transformations enable AI systems to explicitly interpret spatial topology in early design and to access performance indicators essential for performance-driven automated design, optimization, and decision-making.
Two subsets are provided to support both theoretical and practical validation. Dataset 1 consists of procedurally generated, geometry-pure models with limited redundant elements, formatted in B-rep, LPG, and EnergyPlus IDF representations, and is intended for controlled validation of transformation correctness. Dataset 3 contains real-world building models characterized by complex spatial relationships and numerous redundant or non-spatial elements, designed to evaluate the robustness of the cleansing and topology-extraction modules under realistic conditions. The quantitative metrics defined in the manuscript should be applied to assess transformation accuracy, robustness, and efficiency across these datasets.